{"total":13,"offset":0,"limit":20,"items":[{"id":"whisper-local-transcription","kind":"discovery","path":"/discoveries/whisper-local-transcription","title":"Transcribing speech on infrastructure you control","description":"Whisper provides downloadable speech-recognition models for transcription, language identification, translation into English, and caption-making without requiring a hosted speech service.","domains":["Access","Open intelligence"],"published":"2026-09-19","modified":null,"sourceDate":"2022-09-21","dates":{"eventDate":null,"publicationDate":"2022-09-21","captureDate":"2026-09-19","lastReviewedDate":"2026-09-19"},"provenance":{"origin":"editorial","externalId":"https://github.com/openai/whisper"},"revisions":[{"id":"revision:open-models-added:whisper-local-transcription","recordedAt":"2026-09-19","summary":"Bright added this source-checked open-model application record. The cited source publication date is 2022-09-21; 2026-09-19 is when Bright added this record.","sourceIds":["whisper-repository","whisper-release"]}],"corrections":[],"review":"Demonstrated; confidence: unassessed; source-checked; ai-assisted. AI-assisted comparison with the cited sources. Source-checked means the record was checked against those sources; it does not claim independent reproduction, expert review, or validation of the publisher’s results.","sections":[{"title":"The human problem","text":["People need searchable transcripts and captions, including when audio is private or connectivity is limited."]},{"title":"The prior constraint","text":["Strong speech recognition often depended on a remote service or a model tuned to a narrow language and acoustic setting."]},{"title":"AI’s actual role","text":["A multilingual sequence-to-sequence model converts audio into timestamped text or translated text."]},{"title":"The documented result","text":["OpenAI released model weights, inference code, and a command-line interface under MIT terms in September 2022, making local runs broadly reproducible."]},{"title":"Why it may matter","text":["Local weights give organizations more control over where audio travels, but people still need ways to inspect names, meaning, and accessibility quality."]},{"title":"Limitations","text":["Accuracy varies with language, accent, noise, recording conditions, and subject matter. The training corpus and a complete training recipe were not released.","The official repository establishes MIT-licensed code and weights. That unusually permissive artifact release remains distinct from access to the training data."]},{"title":"Unresolved questions","text":[]}],"limitations":["Accuracy varies with language, accent, noise, recording conditions, and subject matter. The training corpus and a complete training recipe were not released.","The official repository establishes MIT-licensed code and weights. That unusually permissive artifact release remains distinct from access to the training data."],"sources":[{"id":"whisper-repository","title":"Whisper","url":"https://github.com/openai/whisper","type":"repository"},{"id":"whisper-release","title":"Introducing Whisper","url":"https://openai.com/index/whisper/","type":"institution"}],"links":[{"title":"Teaching a robot from demonstrations","path":"/discoveries/openvla-robot-manipulation"},{"title":"A shared starting point for humanoid skills","path":"/discoveries/groot-humanoid-skills"},{"title":"Predicting a protein’s shape from its sequence","path":"/discoveries/esmfold-protein-structures"},{"title":"Assembling biological model workflows","path":"/discoveries/bionemo-biological-workflows"},{"title":"Proposing materials for a desired property","path":"/discoveries/mattergen-material-candidates"},{"title":"Open Models","path":"/open-models"},{"title":"Open intelligence","path":"/worlds/open"},{"title":"What changes when powerful models become open-weight?","path":"/threads/open"},{"title":"Someone builds on it","path":"/open-intelligence"}]},{"id":"seamlessm4t-speech-translation","kind":"discovery","path":"/discoveries/seamlessm4t-speech-translation","title":"Moving between speech and text across languages","description":"SeamlessM4T joins speech recognition, text translation, speech translation, and speech generation in one multilingual research model family.","domains":["Access","Open intelligence"],"published":"2026-09-19","modified":null,"sourceDate":"2023-08-22","dates":{"eventDate":null,"publicationDate":"2023-08-22","captureDate":"2026-09-19","lastReviewedDate":"2026-09-19"},"provenance":{"origin":"editorial","externalId":"https://github.com/facebookresearch/seamless_communication"},"revisions":[{"id":"revision:open-models-added:seamlessm4t-speech-translation","recordedAt":"2026-09-19","summary":"Bright added this source-checked open-model application record. The cited source publication date is 2023-08-22; 2026-09-19 is when Bright added this record.","sourceIds":["seamless-communication-repository","seamlessm4t-paper"]}],"corrections":[],"review":"Emerging; confidence: unassessed; source-checked; ai-assisted. AI-assisted comparison with the cited sources. Source-checked means the record was checked against those sources; it does not claim independent reproduction, expert review, or validation of the publisher’s results.","sections":[{"title":"The human problem","text":["Language differences can block conversation and access to spoken information, especially where specialist translation tools are scarce."]},{"title":"The prior constraint","text":["Speech translation pipelines commonly stitched together separate recognition, translation, and synthesis systems."]},{"title":"AI’s actual role","text":["The model accepts speech or text and generates translated speech or text across its documented language set."]},{"title":"The documented result","text":["Meta publishes inference tooling, checkpoints, and supported-language documentation that allow researchers to run and evaluate the model family."]},{"title":"Why it may matter","text":["One inspectable model family can simplify research across the pipeline while making language-by-language failure analysis more urgent."]},{"title":"Limitations","text":["SeamlessM4T v1 and v2 weights are CC-BY-NC-4.0, so commercial use cannot be assumed. Translation quality, toxicity, and speech identity require evaluation for each language and setting.","Capabilities and language coverage are developer-described. MIT code and noncommercial model weights have materially different reuse permissions."]},{"title":"Unresolved questions","text":[]}],"limitations":["SeamlessM4T v1 and v2 weights are CC-BY-NC-4.0, so commercial use cannot be assumed. Translation quality, toxicity, and speech identity require evaluation for each language and setting.","Capabilities and language coverage are developer-described. MIT code and noncommercial model weights have materially different reuse permissions."],"sources":[{"id":"seamless-communication-repository","title":"Seamless Communication","url":"https://github.com/facebookresearch/seamless_communication","type":"repository"},{"id":"seamlessm4t-paper","title":"SeamlessM4T: Massively Multilingual & Multimodal Machine Translation","url":"https://arxiv.org/abs/2308.11596","type":"paper"}],"links":[{"title":"Teaching a robot from demonstrations","path":"/discoveries/openvla-robot-manipulation"},{"title":"A shared starting point for humanoid skills","path":"/discoveries/groot-humanoid-skills"},{"title":"Predicting a protein’s shape from its sequence","path":"/discoveries/esmfold-protein-structures"},{"title":"Assembling biological model workflows","path":"/discoveries/bionemo-biological-workflows"},{"title":"Proposing materials for a desired property","path":"/discoveries/mattergen-material-candidates"},{"title":"Open Models","path":"/open-models"},{"title":"Open intelligence","path":"/worlds/open"},{"title":"What changes when powerful models become open-weight?","path":"/threads/open"},{"title":"Someone builds on it","path":"/open-intelligence"}]},{"id":"bci-speech-gesture-avatar","kind":"discovery","path":"/discoveries/bci-speech-gesture-avatar","title":"One implant, speech and gesture together.","description":"UCSF researchers showed that neural signals recorded through one high-density implant could support simultaneous decoding of attempted speech and communicative gestures in people with paralysis.","domains":["Health","Access"],"published":"2026-09-19","modified":null,"sourceDate":"2026-09-14","dates":{"eventDate":null,"publicationDate":"2026-09-14","captureDate":"2026-09-19","lastReviewedDate":"2026-09-19"},"provenance":{"origin":"editorial","externalId":"https://www.nature.com/articles/s41593-026-02446-2"},"revisions":[{"id":"revision:sept26-ucsf-avatar-01","recordedAt":"2026-09-19","summary":"Initial source-checked record of the three-participant speech-and-gesture proof of concept.","sourceIds":["source-ucsf-avatar-nature"]}],"corrections":[],"review":"Emerging; confidence: high; approved; ai-assisted. AI-assisted editorial comparison with the cited primary sources, explicit evidence limits, and held alternatives. Publication authorized by the site owner on 2026-09-19; no human source review or independent replication is claimed.","sections":[{"title":"The human problem","text":["Paralysis can impair speech and body movement together, narrowing communication to systems that decode only one form of expression at a time."]},{"title":"The prior constraint","text":["Earlier brain–computer interfaces generally decoded speech or gesture in isolation, even though everyday communication often uses both at once."]},{"title":"AI’s actual role","text":["Parallel neural decoders translated high-density electrocorticography signals into selected speech phrases and upper-body gesture commands for a personalized virtual avatar."]},{"title":"The documented result","text":["Researchers decoded isolated movements across three participants and tested the simultaneous full-avatar system with two. Models trained on both isolated and simultaneous attempts performed more reliably across those contexts than models trained on only one context."]},{"title":"Why it may matter","text":["A communication interface that can carry words and gesture at the same time could eventually preserve more of how a person intends to express something, not only which phrase they select."]},{"title":"Limitations","text":["Evidence comes from three participants; the simultaneous full-avatar system was tested with two of them.","The work does not establish long-term home reliability or broad clinical availability.","The system used a wired experimental implant and a small, fixed vocabulary; it did not restore physical movement."]},{"title":"Unresolved questions","text":["How stable is simultaneous control outside the laboratory over months or years?","How can the phrase and gesture vocabulary expand without requiring every combination in training?","Can future systems reduce surgical and specialist-support burdens?"]}],"limitations":["Evidence comes from three participants; the simultaneous full-avatar system was tested with two of them.","The work does not establish long-term home reliability or broad clinical availability.","The system used a wired experimental implant and a small, fixed vocabulary; it did not restore physical movement."],"sources":[{"id":"source-ucsf-avatar-nature","title":"Simultaneous speech and gesture decoding for multimodal communication in paralysis","url":"https://www.nature.com/articles/s41593-026-02446-2","type":"paper"}],"links":[{"title":"A way back to conversation.","path":"/discoveries/speech"},{"title":"A voice in everyday life.","path":"/discoveries/speech-home"},{"title":"Health & biology","path":"/worlds/health"},{"title":"People & culture","path":"/worlds/culture"},{"title":"What if the body could answer intention again?","path":"/threads/agency"}]},{"id":"open-asr-global-south","kind":"discovery","path":"/discoveries/open-asr-global-south","title":"A speech benchmark begins to listen beyond English.","description":"Hugging Face and Voice Arena added Hindi and Indian English evaluation to an open speech-recognition leaderboard with held-out/private splits and demographic and geographic test design.","domains":["Open intelligence","Access"],"published":"2026-09-19","modified":null,"sourceDate":"2026-08-28","dates":{"eventDate":"2026-08-28","publicationDate":"2026-08-28","captureDate":"2026-09-19","lastReviewedDate":"2026-09-19"},"provenance":{"origin":"editorial","externalId":"https://huggingface.co/blog/open-asr-leaderboard-global-south"},"revisions":[{"id":"revision:sept26-asr-01","recordedAt":"2026-09-19","summary":"Initial benchmark-gap draft.","sourceIds":["source-asr-hf"]}],"corrections":[],"review":"Emerging; confidence: medium; approved; ai-assisted. AI-assisted editorial comparison with the cited primary sources, explicit evidence limits, and held alternatives. Publication authorized by the site owner on 2026-09-19; no human source review or independent replication is claimed.","sections":[{"title":"The human problem","text":["Speech tools that look strong on dominant-language benchmarks can fail for accents, languages, and communities missing from evaluation."]},{"title":"The prior constraint","text":["Public speech leaderboards have offered limited Global South language coverage and can be overfit when all test material is visible."]},{"title":"AI’s actual role","text":["Speech-recognition models transcribe the same held-out audio so their errors can be compared across language and population slices."]},{"title":"The documented result","text":["The 28 August launch adds initial Hindi and Indian English coverage and an evaluation design that includes held-out/private splits."]},{"title":"Why it may matter","text":["A better test can expose exclusions before a speech system becomes infrastructure."]},{"title":"Limitations","text":["A benchmark expansion is not proof that any product is equitable.","Two varieties do not represent the Global South.","Final copy must inspect the evaluation assets and demographic documentation."]},{"title":"Unresolved questions","text":["Who consented to and governs the voice data?","Which accents, regions, ages, and environments remain missing?","How will leaderboard gaming and model updates be handled?"]}],"limitations":["A benchmark expansion is not proof that any product is equitable.","Two varieties do not represent the Global South.","Final copy must inspect the evaluation assets and demographic documentation."],"sources":[{"id":"source-asr-hf","title":"The Open ASR Leaderboard Adds Its First Global South Language","url":"https://huggingface.co/blog/open-asr-leaderboard-global-south","type":"institution"}],"links":[{"title":"GoTo’s downloadable AI model makes room for Indonesia’s languages","path":"/discoveries/goto-ai-indonesian-regional-language-model"},{"title":"Open Models","path":"/open-models"},{"title":"Health & biology","path":"/worlds/health"},{"title":"Open intelligence","path":"/worlds/open"},{"title":"People & culture","path":"/worlds/culture"},{"title":"What if the body could answer intention again?","path":"/threads/agency"},{"title":"What changes when powerful models become open-weight?","path":"/threads/open"}]},{"id":"klang-pianissimo-swedish-local-transcription","kind":"discovery","path":"/discoveries/klang-pianissimo-swedish-local-transcription","title":"A Swedish speech model brings fast transcription closer to home.","description":"Klang’s released speech model offers a practical choice about where Swedish audio is processed. Its benchmark gains are promising, with important differences between speed and accuracy.","domains":["Access","Work","Open intelligence"],"published":"2026-10-16","modified":null,"sourceDate":"2026-09-23","dates":{"captureDate":"2026-10-16","eventDate":"2026-09-23","lastReviewedDate":"2026-10-16","publicationDate":"2026-09-23"},"provenance":{"origin":"editorial","externalId":"https://klang.ai/pianissimo/"},"revisions":[{"id":"revision:klang-pianissimo-swedish-local-transcription:initial","recordedAt":"2026-10-16","sourceIds":["source:klang-pianissimo-swedish-local-transcription:1","source:klang-pianissimo-swedish-local-transcription:2","source:klang-pianissimo-swedish-local-transcription:3","source:klang-pianissimo-swedish-local-transcription:4","source:klang-pianissimo-swedish-local-transcription:5","source:klang-pianissimo-swedish-local-transcription:6","source:klang-pianissimo-swedish-local-transcription:7","source:klang-pianissimo-swedish-local-transcription:8","source:klang-pianissimo-swedish-local-transcription:9"],"summary":"Initial source-checked account of the released Swedish speech model, benchmark comparisons, local and demo data routes, setup requirements and limits."}],"corrections":[],"review":"Emerging; confidence: unassessed; source-checked; ai-assisted. Source-audited manuscript supplied October 16. September release dates remain distinct from the October 14 NVIDIA roundup. Company benchmarks are not independent reproduction; no runtime test by Bright is claimed.","sections":[{"title":"Editorial image","text":["A microphone on a small tripod beside a laptop displaying audio-mixing software, with a portable recorder on the desk.: https://brightaifuture.com/media/content/fc882e0c90090f0596a06288a8b564eeeb55aba88d44c90a0eb04519501a5f52.jpg","Credit: A microphone and laptop set up for audio recording. Context photo, not a demonstration of Pianissimo. Photo: Jeremy Enns / Pexels.. License: Pexels License — free website use; https://www.pexels.com/license/.","Source: https://www.pexels.com/photo/a-laptop-beside-a-microphone-in-a-studio-5083606/","contextual; not Pianissimo software, a Klang product demonstration, benchmark evidence or endorsement."]},{"title":"The story","text":["Turning a recording into text can make an interview easier to search, a lecture easier to revisit or a draft easier to write. For Swedish speakers, a newly released speech model offers another way to do that work on their own computers.","Klang’s Pianissimo adapts NVIDIA’s Parakeet v3 for Swedish. Its downloadable model weights were released on September 23, followed by smaller local-computer versions on September 28. NVIDIA highlighted it in an October 14 roundup of Nordic AI projects. This is software people can obtain now, although its reported performance still needs testing in the situations where they would use it. [Klang’s release page](https://klang.ai/pianissimo/) · [NVIDIA’s roundup](https://blogs.nvidia.com/blog/nordics-technology-ecosystem/)","The model has about 600 million parameters and is released under CC BY 4.0. It produces text with punctuation and can supply word-level timestamps. Klang says it fine-tuned the system on approximately 50,000 hours of Swedish speech. [Model documentation](https://huggingface.co/KlangAI/pianissimo-sv)"]},{"title":"Two gains, two different comparisons","text":["The attractive headline numbers use different baselines: roughly 75% fewer transcription errors than original Parakeet, and about 64 times the bulk processing speed of KB-Whisper large.","In Klang’s evaluation, which gave equal weight to five datasets, mean word error rates were 17.7% for Parakeet, 4.5% for Pianissimo and 3.5% for KB-Whisper large. Lower is better, so KB-Whisper large remained more accurate on that average.","On an NVIDIA A100 graphics processor, Pianissimo processed audio at roughly 2,500 times real time, against 39 times for KB-Whisper large. Klang tested 30-second clips in batches, using different software frameworks and each model’s fastest tested configuration, excluding model loading. That does not establish a 64-fold speed-up on an ordinary laptop. [Klang’s technical report](https://research.klang.ai/papers/how-we-trained-pianissimo)","That trade-off matters. Someone processing a large archive may value throughput differently from someone checking a short, difficult interview. The useful question is how quickly they can get a transcript they trust, including the time spent correcting it."]},{"title":"What running locally changes","text":["Klang has released versions that run on ordinary computer processors, including a 660 MB version, alongside versions for Apple-silicon graphics hardware. Their published speed tests name the hardware used, so readers can separate availability from the performance of a particular machine. [CPU version](https://huggingface.co/KlangAI/pianissimo-sv-onnx) · [Mac version](https://huggingface.co/KlangAI/pianissimo-sv-mlx)","The company also provides a command-line tool for local files. Its documentation says transcription runs on the computer, with no audio upload to a transcription service, and local files can be processed offline after setup. This particular tool downloads the full 2.51 GB model, rather than the smaller versions above, and its documentation asks for at least 10 GB of free space for a CPU installation. It also requires Node.js, Python and FFmpeg. This route will be more approachable for people comfortable with technical setup. [Pianissimo CLI](https://github.com/klang-ai/pianissimo-cli)","There is an important distinction when trying it: the demonstration on Klang’s website sends uploaded audio to Klang. Downloading a model for local use and uploading a recording to a web demonstration are different choices about where that recording goes. [Demo disclosure](https://klang.ai/pianissimo/)","Keeping processing on a computer can reduce the need to send recordings elsewhere. It does not settle every privacy question. The CLI documentation says completed transcripts are cached locally. People still need to consider who can access their device, how recordings and transcripts are stored, and whether everyone involved is comfortable being recorded. [Storage documentation](https://github.com/klang-ai/pianissimo-cli#storage)"]},{"title":"The next test is everyday speech","text":["Klang warns that overlapping voices, heavy background noise, unusual names and specialist vocabulary can reduce recognition quality. Performance on other languages and switching between languages has not been established. [Model limitations](https://huggingface.co/KlangAI/pianissimo-sv#limitations)","Bright’s assessment is that this release is worth following because it combines language-specific development with a concrete choice of where the software runs. A useful next test would compare the total effort needed to produce checked transcripts from real interviews, meetings and lectures, including difficult passages.","The evidence so far supports a released tool and encouraging company benchmarks. Whether it saves people time, improves access to information or handles their particular conversations well remains a question for use and independent evaluation."]},{"title":"The human problem","text":["Swedish speakers need searchable, checkable transcripts and a practical choice about where recordings are processed."]},{"title":"The prior constraint","text":["Transcription options differ in accuracy, processing speed, setup requirements and whether audio leaves the device."]},{"title":"AI’s actual role","text":["Pianissimo adapts NVIDIA Parakeet v3 to recognize Swedish speech, with local CPU and Apple-silicon runtimes."]},{"title":"The documented result","text":["Weights were released September 23, 2026, with local variants on September 28; NVIDIA highlighted the release October 14. Klang reports mean word error rates of 17.7% for Parakeet, 4.5% for Pianissimo and 3.5% for KB-Whisper large over five equally weighted datasets, and A100 batch processing at roughly 2,500 times real time versus 39 times for KB-Whisper large."]},{"title":"Why it may matter","text":["A language-specific released model provides a concrete local-processing option for people producing Swedish transcripts, while practical value still depends on setup and correction effort."]},{"title":"Limitations","text":["Results are company-reported. Bright has not run the software, reproduced benchmarks or independently audited its data flows.","The accuracy and speed headline figures compare different baselines; KB-Whisper large remained more accurate on the reported mean.","A100 throughput on batched 30-second clips, excluding model loading and using different frameworks, does not establish laptop speed.","The 660 MB ONNX version is separate from the CLI’s full 2.51 GB model download and documented 10 GB free-space CPU requirement.","The website demo uploads audio to Klang; local runtimes and hosted services have different data paths.","Completed CLI transcripts are cached locally; local processing alone does not establish privacy.","Overlapping voices, heavy noise, unusual names and specialist vocabulary can reduce quality. Other-language and code-switching performance is not established.","Time savings, accessibility gains and everyday usefulness remain unmeasured."]},{"title":"Unresolved questions","text":["How much total effort, including transcript correction and technical setup, does Pianissimo save on real Swedish interviews, meetings and lectures?"]}],"limitations":["Results are company-reported. Bright has not run the software, reproduced benchmarks or independently audited its data flows.","The accuracy and speed headline figures compare different baselines; KB-Whisper large remained more accurate on the reported mean.","A100 throughput on batched 30-second clips, excluding model loading and using different frameworks, does not establish laptop speed.","The 660 MB ONNX version is separate from the CLI’s full 2.51 GB model download and documented 10 GB free-space CPU requirement.","The website demo uploads audio to Klang; local runtimes and hosted services have different data paths.","Completed CLI transcripts are cached locally; local processing alone does not establish privacy.","Overlapping voices, heavy noise, unusual names and specialist vocabulary can reduce quality. Other-language and code-switching performance is not established.","Time savings, accessibility gains and everyday usefulness remain unmeasured."],"sources":[{"id":"source:klang-pianissimo-swedish-local-transcription:1","title":"Demo disclosure","type":"institution","url":"https://klang.ai/pianissimo/"},{"id":"source:klang-pianissimo-swedish-local-transcription:2","title":"NVIDIA’s roundup","type":"institution","url":"https://blogs.nvidia.com/blog/nordics-technology-ecosystem/"},{"id":"source:klang-pianissimo-swedish-local-transcription:3","title":"Model documentation","type":"repository","url":"https://huggingface.co/KlangAI/pianissimo-sv"},{"id":"source:klang-pianissimo-swedish-local-transcription:4","title":"Klang’s technical report","type":"report","url":"https://research.klang.ai/papers/how-we-trained-pianissimo"},{"id":"source:klang-pianissimo-swedish-local-transcription:5","title":"CPU version","type":"repository","url":"https://huggingface.co/KlangAI/pianissimo-sv-onnx"},{"id":"source:klang-pianissimo-swedish-local-transcription:6","title":"Mac version","type":"repository","url":"https://huggingface.co/KlangAI/pianissimo-sv-mlx"},{"id":"source:klang-pianissimo-swedish-local-transcription:7","title":"Pianissimo CLI","type":"repository","url":"https://github.com/klang-ai/pianissimo-cli"},{"id":"source:klang-pianissimo-swedish-local-transcription:8","title":"Storage documentation","type":"repository","url":"https://github.com/klang-ai/pianissimo-cli#storage"},{"id":"source:klang-pianissimo-swedish-local-transcription:9","title":"Model limitations","type":"repository","url":"https://huggingface.co/KlangAI/pianissimo-sv#limitations"}],"links":[]},{"id":"open-model-voxtral","kind":"model-family","path":"/open-models/voxtral","title":"Voxtral Realtime: license, size and what is open","description":"A compact streaming speech-recognition model designed for low-latency transcription through a realtime API.","domains":["Open intelligence"],"published":"2026-09-19","modified":"2026-10-11","sourceDate":null,"review":"AI-assisted primary-source check on 2026-09-19; no independent model reproduction or license certification. vLLM integration corroborates production serving support; transcription quality claims are not independently reproduced here.","sections":[{"title":"Answer at a glance","text":["Commercial use permitted under Apache-2.0; review accompanying terms.","About 8.86 GB of selected documented weights","Selected artifact: mistralai/Voxtral-Mini-4B-Realtime-2602. Format: Safetensors, Mistral consolidated representation. File evidence checked 2026-10-11. File listing: https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602/tree/2769294da9567371363522aac9bbcfdd19447add.","Single consolidated.safetensors. Excludes alternative model.safetensors (8,859,446,848 bytes); alternate representations are not complementary shards. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.","Runtime memory, compatibility and speed on your computer are untested by Bright. Inspect documented candidates at https://brightaifuture.com/open-models#hardware."]},{"title":"Documented release","text":["Voxtral-Mini-4B-Realtime-2602","Organization: Mistral AI","Voxtral Realtime release: 2026-02-04 (day precision). Source: https://mistral.ai/news/voxtral-transcribe-2/","Parameters: 4B","Architecture: Streaming audio encoder-decoder model integrated with Mistral inference runtimes.","Modalities: audio, text","Context: Streaming sessions; a comparable language-token window is not asserted here."]},{"title":"License and commercial use","text":["Apache-2.0","Permitted by Apache-2.0."]},{"title":"What is open","text":["Weights: available. Official weights are downloadable. Source: https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602","Architecture: available. Configuration and model card are public. Source: https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602","Inference code: available. The release documents vLLM’s Realtime API integration. Source: https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602","Training code: not-established. Complete training code was not confirmed. Source: https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602","Training recipe: not-established. A full reproducible training recipe was not confirmed. Source: https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602","Data information: partial. Capabilities and languages are described without a complete corpus inventory. Source: https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602","Training data: not-established. No sufficiently specific public artifact was confirmed in this review. Source: https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602","Evaluation: partial. Vendor speech evaluations are published in the model materials. Source: https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602","Commercial use: available. The official checkpoint is Apache-2.0 licensed. Source: https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602"]},{"title":"Uses and strengths described in sources","text":["Streaming transcription","Low-latency audio processing","Compact size"]},{"title":"Hardware and quantization","text":["The 4B size supports comparatively modest deployment, but streaming overhead and audio session length affect requirements.","Runtime-supported formats should be verified against the official card"]},{"title":"Independent evidence","text":["vLLM integration corroborates production serving support; transcription quality claims are not independently reproduced here."]},{"title":"Limitations","text":["Speech-focused rather than a general multimodal assistant","Training data is not fully disclosed"]}],"sources":[{"id":"voxtral-rt-release","title":"Voxtral Transcribe 2 and Realtime release","url":"https://mistral.ai/news/voxtral-transcribe-2/","type":"institution"},{"id":"voxtral-rt-card","title":"Voxtral Mini 4B Realtime model card","url":"https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602","type":"registry"},{"id":"mistral-changelog","title":"Mistral model changelog","url":"https://docs.mistral.ai/resources/changelogs","type":"institution"}],"limitations":["Speech-focused rather than a general multimodal assistant","Training data is not fully disclosed"],"links":[{"title":"Open Models","path":"/open-models"},{"title":"Open intelligence thread","path":"/threads/open"}]},{"id":"speech","kind":"discovery","path":"/discoveries/speech","title":"A way back to conversation.","description":"A man with ALS used an experimental brain–computer interface to turn attempted speech into conversation.","domains":["Access"],"published":"2026-09-05","modified":null,"sourceDate":"2024-08-14","dates":{"eventDate":null,"publicationDate":"2024-08-14","captureDate":null,"lastReviewedDate":"2026-09-05"},"provenance":{"origin":"legacy-projection","externalId":"speech"},"revisions":[],"corrections":[],"review":"Emerging; confidence: unassessed; source-checked; legacy source check. Legacy source check; no named human reviewer is recorded in this projection.","sections":[{"title":"The human problem","text":["ALS can leave thoughts intact while taking away the ability to speak clearly."]},{"title":"The prior constraint","text":["Communication depended on limited movement and other people interpreting speech."]},{"title":"AI’s actual role","text":["Neural decoders translated recorded cortical activity into words, then synthesized speech."]},{"title":"The documented result","text":["The study reported 97.5% word accuracy after further training, sustained over 8.4 months."]},{"title":"Why it may matter","text":["For one participant, the system supported self-paced conversations."]},{"title":"Limitations","text":["One participant, implanted electrodes, and specialist support. This is not a broadly available treatment."]},{"title":"Unresolved questions","text":["Test reliability, safety, and usefulness with more participants and longer follow-up."]}],"limitations":["One participant, implanted electrodes, and specialist support. This is not a broadly available treatment."],"sources":[{"id":"source:speech","title":"An Accurate and Rapidly Calibrating Speech Neuroprosthesis · NEJM","url":"https://www.nejm.org/doi/full/10.1056/NEJMoa2314132","type":"paper"}],"links":[{"title":"Reconnecting intention and movement.","path":"/discoveries/spine"},{"title":"Helping specialists see urgency.","path":"/discoveries/retina"},{"title":"Health & biology","path":"/worlds/health"},{"title":"People & culture","path":"/worlds/culture"},{"title":"What if the body could answer intention again?","path":"/threads/agency"},{"title":"Extended story and historical record","path":"/stories/a-way-back-to-conversation"},{"title":"Open this record in the interactive world","path":"/?event=speech"}]},{"id":"legato-hearing-glasses","kind":"discovery","path":"/discoveries/legato-hearing-glasses","title":"Legato puts hearing assistance in eyeglass frames. The tradeoffs matter.","description":"Legato says its U.S. launch pairs on-frame speech processing with open-ear audio and prescription-lens options. Listed prices start at $999; listening, fit and availability will determine its usefulness.","domains":["Access","Health"],"published":"2026-10-02","modified":null,"sourceDate":"2026-10-01","dates":{"eventDate":"2026-10-01","publicationDate":"2026-10-01","captureDate":"2026-10-02","lastReviewedDate":"2026-10-02"},"provenance":{"origin":"editorial","externalId":"https://www.globenewswire.com/de/news-release/2026/10/01/3372968/0/en/legato-frames-the-world-s-first-ai-based-hearing-glasses-are-now-available-in-the-u-s.html"},"revisions":[{"id":"revision:012c9bf1e8ecd6d13ab0","recordedAt":"2026-10-02","summary":"Legato puts hearing assistance in eyeglass frames. The tradeoffs matter.","sourceIds":["source:legato-us-launch","source:legato-how-it-works","source:legato-symphony","source:legato-harmony","source:legato-aria","source:fda-otc-hearing-category","source:hearadvisor-method"]}],"corrections":[],"review":"Emerging; confidence: unassessed; approved; ai-assisted. Owner-authorized AI-assisted source comparison, with parent editorial review and an October 2 default-configuration purchase-interface check. Manufacturer specifications and privacy claims are attributed. FDA category context is distinct from product approval. No hands-on test, independent clinical outcome, privacy audit, order or delivery verification is claimed.","sections":[{"title":"The human problem","text":["A conversation is more than audible sound. Following a colleague in a busy room can require effort, and a useful device must fit into the life of the person wearing it. Familiar eyeglass frames offer one possible route to hearing assistance without an object in the ear canal."]},{"title":"The prior constraint","text":["Choosing hearing assistance involves comfort, adjustment, cost and daily use as well as sound. Combining it with eyewear changes those choices: lens needs, frame fit and hearing support become part of the same purchase."]},{"title":"AI’s actual role","text":["Legato says SpeechSense uses an on-device neural network to separate speech from background noise and enhance it in real time. Open-ear speakers deliver the sound. The company says processing happens on the frames, with no camera or audio recording. These are manufacturer descriptions, not a Bright privacy audit or measured listening result."]},{"title":"The documented result","text":["Legato announced its U.S. launch on October 1 at 09:00 Eastern. Its product pages checked October 2 list Symphony and Harmony at $999 and Aria at $1,099, before any separately priced lens options. The frames support prescription lenses and Bluetooth music and calls on iOS and Android. Product pages advertise more than ten hours of power; that is a manufacturer specification. The announcement establishes a product launch, not an observed improvement in everyday conversations.\n\nOn October 2, between 18:39:42 and 18:39:49 Eastern, Bright rechecked Symphony Narrow/Matte Black, Harmony Medium/Matte Black and Aria Medium/Classic Tortoise. Each default product page displayed “Select Lenses.” After choosing nonprescription lenses with the free clear basic hardcoat, each review screen displayed “Add to Cart”; no “Sold Out” or “Notify me” message was visible in those checked flows. These are interface observations, not confirmed inventory or delivery. Bright did not add an item to a cart or place an order, and the check does not cover other variants or retailers."]},{"title":"Why it may matter","text":["Bright’s analysis: the practical possibility is a familiar object that helps someone stay in a conversation. For a person who already wears glasses, one device could reduce the friction of carrying and maintaining separate equipment. For someone who does not, wearing frames may introduce a new obligation. Neither response can be inferred from a launch.\n\nThe useful test is ordinary life: a quiet conversation, several people around a table, a noisy workplace and a full day of wearing the frames. Does speech become easier to follow? Is the sound comfortable, and can nearby people hear it? Fit, lens quality, battery behavior and reliable support matter alongside the processing. Bright has not performed these tests.\n\nCost and access are part of the result. Compare the complete price for the chosen frame and lenses, the return and warranty terms, and the availability of adjustment and hearing support. A starting price does not establish affordability or insurance coverage.\n\nFDA’s OTC category is for adults 18 and older with perceived mild to moderate hearing loss; it is not the category for children or severe or profound hearing loss. Legato describes its frames as OTC hearing aids. That description is not a claim here of FDA approval or clearance.\n\nThe next evidence should describe who used the frames, in what listening conditions, with what comparison and for how long. Independent listening results, sustained use and accessible support would make the possibility easier to judge."]},{"title":"Limitations","text":["Emerging evidence: the checked material establishes an announced product and manufacturer specifications. Bright has no hands-on evaluation, clinical outcome study or observed long-term user result.","Legato cites HearAdvisor testing. Bright checked the lab’s general measurement method, but did not inspect a product-specific report supporting the advertised score or comparisons, so no score or efficacy ranking is repeated. Laboratory metrics cannot establish an individual wearer’s benefit.","On-device processing, no recording and battery life remain company claims. Bright has not audited the device or app, measured sound leakage or battery endurance, or verified every purchase configuration, delivery estimate, lens price or insurance benefit."]},{"title":"Unresolved questions","text":["Who benefits in everyday speech-in-noise conditions, and how do independent comparisons account for fit and hearing needs?","What are the complete lens and support costs, and which configurations can actually be delivered?","Do comfort, battery life and sustained use hold outside laboratory and company demonstrations?"]}],"limitations":["Emerging evidence: the checked material establishes an announced product and manufacturer specifications. Bright has no hands-on evaluation, clinical outcome study or observed long-term user result.","Legato cites HearAdvisor testing. Bright checked the lab’s general measurement method, but did not inspect a product-specific report supporting the advertised score or comparisons, so no score or efficacy ranking is repeated. Laboratory metrics cannot establish an individual wearer’s benefit.","On-device processing, no recording and battery life remain company claims. Bright has not audited the device or app, measured sound leakage or battery endurance, or verified every purchase configuration, delivery estimate, lens price or insurance benefit."],"sources":[{"id":"source:legato-us-launch","title":"Legato U.S. launch · October 1, 2026, 09:00 ET · manufacturer announcement","url":"https://www.globenewswire.com/de/news-release/2026/10/01/3372968/0/en/legato-frames-the-world-s-first-ai-based-hearing-glasses-are-now-available-in-the-u-s.html","type":"institution"},{"id":"source:legato-how-it-works","title":"Legato mechanism and battery claims · checked October 2, 2026","url":"https://www.legatohearing.com/pages/how-it-works","type":"institution"},{"id":"source:legato-symphony","title":"Symphony product · $999 listed frame price · checked October 2, 2026","url":"https://www.legatohearing.com/products/symphony","type":"institution"},{"id":"source:legato-harmony","title":"Harmony product and frame fit guide · $999 listed frame price · checked October 2, 2026","url":"https://www.legatohearing.com/products/harmony","type":"institution"},{"id":"source:legato-aria","title":"Aria product · $1,099 listed frame price · checked October 2, 2026","url":"https://www.legatohearing.com/products/aria","type":"institution"},{"id":"source:fda-otc-hearing-category","title":"FDA OTC hearing aid category · adult eligibility and limits","url":"https://www.fda.gov/medical-devices/hearing-aids/otc-hearing-aids-what-you-should-know","type":"government"},{"id":"source:hearadvisor-method","title":"HearAdvisor measurement method · laboratory comparisons, not an individual outcome","url":"https://www.hearadvisor.com/how-we-measure","type":"institution"}],"links":[{"title":"Health & biology","path":"/worlds/health"},{"title":"People & culture","path":"/worlds/culture"},{"title":"What if the body could answer intention again?","path":"/threads/agency"}]},{"id":"speech-home","kind":"discovery","path":"/discoveries/speech-home","title":"A voice in everyday life.","description":"Casey Harrell used a brain–computer interface at home to communicate and operate his computer without researchers present.","domains":["Access"],"published":"2026-09-06","modified":null,"sourceDate":"2026-06-15","dates":{"eventDate":null,"publicationDate":"2026-06-15","captureDate":null,"lastReviewedDate":"2026-09-06"},"provenance":{"origin":"legacy-projection","externalId":"speech-home"},"revisions":[],"corrections":[],"review":"Emerging; confidence: unassessed; source-checked; legacy source check. Abstract and figure captions checked via PubMed; corroborated with UC Davis and NIH accounts. Full paper not reviewed by Bright.","sections":[{"title":"The human problem","text":["A communication tool must work in daily life, beyond supervised research sessions."]},{"title":"The prior constraint","text":["Earlier use required researchers to be present. Independence and long-term reliability remained open questions."]},{"title":"AI’s actual role","text":["Neural decoders translated attempted speech and movement into text and computer cursor control."]},{"title":"The documented result","text":["The study reported more than 3,800 hours of home use by one participant, including communication and computer access, without researchers present."]},{"title":"Why it may matter","text":["Harrell could communicate with family and friends and use his own computer."]},{"title":"Limitations","text":["One participant with an implanted investigational device. Caregivers still helped set up the system. This is not general clinical availability or independent replication."]},{"title":"Unresolved questions","text":["Test safety and reliability with more participants; reduce the equipment and support needed for daily use."]}],"limitations":["One participant with an implanted investigational device. Caregivers still helped set up the system. This is not general clinical availability or independent replication."],"sources":[{"id":"source:speech-home","title":"Long-term independent use of an intracortical brain–computer interface for speech and cursor control · Nature Medicine (PubMed)","url":"https://pubmed.ncbi.nlm.nih.gov/42297978/","type":"paper"}],"links":[{"title":"A way back to conversation.","path":"/discoveries/speech"},{"title":"Reconnecting intention and movement.","path":"/discoveries/spine"},{"title":"Health & biology","path":"/worlds/health"},{"title":"People & culture","path":"/worlds/culture"},{"title":"What if the body could answer intention again?","path":"/threads/agency"},{"title":"Extended story and historical record","path":"/stories/a-voice-in-everyday-life"},{"title":"Open this record in the interactive world","path":"/?event=speech-home"}]},{"id":"singapore-on-device-ai-scam-warning","kind":"discovery","path":"/discoveries/singapore-on-device-ai-scam-warning","title":"An AI scam warning before you send the money","description":"Google DeepMind and Singapore’s GovTech are developing on-device scam protection. The goal is a timely warning while keeping conversations private. Public rollout is still ahead.","domains":["Public services","Computing"],"published":"2026-10-19","modified":null,"sourceDate":"2026-10-19","dates":{"captureDate":"2026-10-19","eventDate":"2026-10-14","lastReviewedDate":"2026-10-19","publicationDate":"2026-10-19"},"provenance":{"origin":"editorial","externalId":"https://deepmind.google/blog/on-device-ai-scam-detection-mddi/"},"revisions":[{"id":"singapore-on-device-ai-scam-warning-initial","recordedAt":"2026-10-19","sourceIds":["source:singapore-on-device-ai-scam-warning:government","source:singapore-on-device-ai-scam-warning:technical","source:singapore-on-device-ai-scam-warning:scamshield"],"summary":"Initial owner-approved source-based article, preserving public-rollout and effectiveness limitations."}],"corrections":[],"review":"Emerging; confidence: unassessed; approved; ai-assisted. Owner-approved reviewed article supplied for publication. Developer announcement corroborated by a government statement; no published real-world effectiveness results in the sources reviewed.","sections":[{"title":"An AI scam warning before you send the money","text":["Imagine a message arriving from someone who seems to know exactly how to win your trust. The conversation feels ordinary until it turns into a request for money. A useful moment for help would be just before you act, while there is still time to stop and check.","That is the opportunity behind a collaboration between Google DeepMind and Singapore’s Government Technology Agency, GovTech. At Singapore International Cyber Week on October 14, minister Josephine Teo confirmed that the partners are developing real-time scam detection on people’s devices. The aim is to identify manipulation early while keeping personal messages on the device. [Singapore government announcement](https://www.mddi.gov.sg/newsroom/opening-address-by-minister-josephine-teo-at-the-singapore-international-cyber-week--day-2-/)"]},{"title":"Help inside the conversation","text":["Google describes a two-stage system: a lightweight message check, followed by closer conversation analysis using a Gemma model fine-tuned on local scam examples. Suspicious patterns trigger a warning. [Technical announcement](https://deepmind.google/blog/on-device-ai-scam-detection-mddi/)","For a person caught in a convincing exchange, that could be valuable breathing room. Bright’s view is that the appeal lies in the timing: a warning can create an opportunity to reconsider, rather than simply explain what went wrong afterward. Whether people notice, trust and act on it is part of what a real-world rollout needs to establish."]},{"title":"Keeping the check close to you","text":["Google says processing stays on-device, with consent governed by the host app’s terms. Neither Google nor the government would receive conversations through this deployment. [Privacy design](https://deepmind.google/blog/on-device-ai-scam-detection-mddi/)","That is an important design choice for a tool handling private conversations. Readers should still examine the permissions and privacy information of any eventual app. The model’s processing location describes this check, not every other feature an app might offer."]},{"title":"One more place to interrupt a scam","text":["In her speech, Teo described a scam moving from social media to messaging, then a phone call and finally a financial institution. Different organizations see different pieces. The implication is practical: a useful new warning would complement other defenses, rather than make the whole chain someone else’s problem. [Ministerial speech](https://www.mddi.gov.sg/newsroom/opening-address-by-minister-josephine-teo-at-the-singapore-international-cyber-week--day-2-/)","Singapore already has consumer-facing tools. A May government statement describes ScamShield’s ability to block verified scam calls, filter scam SMS messages, and let people check or report suspicious calls, messages and links. That existing service should not be confused with an announced rollout of this new model. [ScamShield’s existing functions](https://www.mddi.gov.sg/newsroom/mddi-s-response-to-pq-on-accuracy-benchmarks-and-liability-frameworks-for-intelligence-deepfake-detector-before-public-rollout-and-integrating-real-time-media-verification-into-scamshield/)"]},{"title":"What comes next","text":["Google plans phased app deployment. Its announcement gives no launch date, named consumer app or detection and false-alarm rates. As of October 19, this remains a development to follow, rather than an announced tool readers can install. [Rollout status](https://deepmind.google/blog/on-device-ai-scam-detection-mddi/)","The next useful evidence would show how often warnings catch genuine manipulation, how often they interrupt harmless conversations, and whether they help people avoid losses in everyday use.","The encouraging direction is clear: AI designed to give people a little more time and agency at a vulnerable moment. Turning that design into dependable protection is the next step."]},{"title":"The human problem","text":["Scammers can manipulate people during private conversations into sending money."]},{"title":"The prior constraint","text":["Different organizations see different parts of a scam, and intervention may arrive after someone acts."]},{"title":"AI’s actual role","text":["A lightweight message check and a locally fine-tuned Gemma model analyze conversations on-device and warn about suspicious patterns."]},{"title":"The documented result","text":["Google DeepMind and GovTech announced development of real-time on-device scam detection on October 14, 2026, corroborated by a Singapore government statement."]},{"title":"Why it may matter","text":["A timely warning could give people more time to reconsider before sending money."]},{"title":"Limitations","text":["Developer announcement corroborated by a government statement. No published real-world effectiveness results in the sources reviewed.","Public rollout is still ahead. No launch date, named consumer app, detection rates or false-alarm rates are given.","Processing location describes this check, not every other feature an eventual host app may offer."]},{"title":"Unresolved questions","text":["How often will warnings identify genuine manipulation, interrupt harmless conversations, and help people avoid losses in everyday use?"]}],"limitations":["Developer announcement corroborated by a government statement. No published real-world effectiveness results in the sources reviewed.","Public rollout is still ahead. No launch date, named consumer app, detection rates or false-alarm rates are given.","Processing location describes this check, not every other feature an eventual host app may offer."],"sources":[{"id":"source:singapore-on-device-ai-scam-warning:government","title":"Singapore government announcement","type":"government","url":"https://www.mddi.gov.sg/newsroom/opening-address-by-minister-josephine-teo-at-the-singapore-international-cyber-week--day-2-/"},{"id":"source:singapore-on-device-ai-scam-warning:technical","title":"Google DeepMind technical announcement","type":"institution","url":"https://deepmind.google/blog/on-device-ai-scam-detection-mddi/"},{"id":"source:singapore-on-device-ai-scam-warning:scamshield","title":"ScamShield’s existing functions","type":"government","url":"https://www.mddi.gov.sg/newsroom/mddi-s-response-to-pq-on-accuracy-benchmarks-and-liability-frameworks-for-intelligence-deepfake-detector-before-public-rollout-and-integrating-real-time-media-verification-into-scamshield/"}],"links":[]},{"id":"google-genie-4-interactive-worlds-audio","kind":"discovery","path":"/discoveries/google-genie-4-interactive-worlds-audio","title":"Google’s Genie 4 adds sound and conversation to worlds you can explore","description":"Google DeepMind’s new world model lets testers steer generated scenes while their surroundings and characters respond. Access remains limited, and a convincing scene is not a validated simulation.","domains":["Computing","Culture","Robotics"],"published":"2026-10-28","modified":null,"sourceDate":"2026-10-27","dates":{"captureDate":"2026-10-28","eventDate":"2026-10-27","lastReviewedDate":"2026-10-28","publicationDate":"2026-10-27"},"provenance":{"origin":"editorial","externalId":"https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/"},"revisions":[{"id":"google-genie-4-interactive-worlds-audio-initial","recordedAt":"2026-10-28","sourceIds":["source:google-genie-4-interactive-worlds-audio:1","source:google-genie-4-interactive-worlds-audio:2"],"summary":"Added source-checked announcement reporting, practical stakes, exact access and licensing caveats, tags and original editorial cover."}],"corrections":[],"review":"Experimental; confidence: unassessed; approved; ai-assisted. Primary announcement and official model/access documentation source-checked October 28. Company claims attributed. No independent hands-on tests. Original AI-generated illustration inspected, explicitly credited and distinguished from model output.","sections":[{"title":"Editorial image","text":["Editorial illustration of a person exploring a miniature landscape unfolding from a monitor, with sound waves and speech shapes connecting its characters.: https://brightaifuture.com/media/content/0ae91d74b5a2bf12d5c30db598380e941b47e4ac68cd51bd1350f76d38f2582f.png","Credit: AI-generated editorial illustration for Bright; not model output. License: Original artwork commissioned for Bright.","Source: https://brightaifuture.com/media/content/0ae91d74b5a2bf12d5c30db598380e941b47e4ac68cd51bd1350f76d38f2582f.png","illustration; not a Genie 4 screenshot, actual model output, a real location or validated simulation."]},{"title":"Google’s Genie 4 adds sound and conversation to worlds you can explore","text":["Google DeepMind announced Genie 4 on October 27, adding native audio and interactive dialogue to its real-time world model. Instead of watching a finished video, a tester can navigate a generated environment and change it as it unfolds. Google says sounds, music and speech are generated alongside the scene, with characters responding to the user and events around them. [Google’s announcement](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/)","That moves the experiment closer to an explorable sketch: a place whose appearance, camera movement, atmosphere and conversations can be tried together. Google’s published demonstrations show what the company has selected to present. Bright has not independently tested Genie 4 or measured how consistently it follows instructions."]},{"title":"A scene that responds while you are inside it","text":["Google describes keyboard controls for movement and camera control, with text prompts used to alter the developing scene. A tester can ask for different weather, introduce an object or person, or change the camera. Native sound adds another layer: footsteps, environmental noise and characters’ voices can accompany the generated world. These are company-described capabilities, not evidence that every requested interaction works reliably. [Announcement and demonstrations](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/)","The system also draws on Google’s Gemini models. Google says Gemini Omni helps anchor appearance and behavior, while a guiding Gemini model observes incoming frames, breaks goals into steps and adapts its plan. This is a combination of visual generation and model-guided planning. It does not establish that the resulting world obeys real physics or represents a real place accurately."]},{"title":"A possible new draft space for creative work","text":["For a filmmaker, the useful possibility is trying the feel of a scene before committing to production. A location could be explored from different viewpoints while lighting, action and dialogue change. Google says trusted testers are investigating this kind of previsualization, alongside game concept design and other scenario-generation work. Those are reported experiments, rather than demonstrated industry-wide savings or finished production workflows. [Trusted tester examples](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/)","Our reading is that the immediate promise lies in helping people decide what to make. A short, responsive draft may help a team communicate a visual idea or notice that a scene does not work. Whether that advantage survives inconsistent details, revision demands and the cost of running the system remains an open question. A prototype still needs a route into the tools people use to finish and deliver their work.","Google also describes early robotics research involving Genie and Gemini. Generated environments might eventually help agents encounter more varied situations. That prospect needs a stricter test than visual plausibility: a world used for training or evaluation must be assessed for how its errors affect what an agent learns. Google’s announcement does not establish reliable transfer from these demonstrations to physical robots."]},{"title":"The limits affect what the worlds are useful for","text":["Google acknowledges that visual details and object permanence can drift, especially during longer sessions, complex edits or navigation. Complex physical interactions, spatial accuracy, legible text, audio accuracy and long-term consistency remain research problems. Its model page says continuous interaction lasts a few minutes, rather than extended sessions. [Google’s limitations](https://deepmind.google/models/genie/)","These limits matter in practice. A prop that changes when the camera turns can undermine continuity planning. A plausible flood scene cannot, by itself, predict how water would move through a particular neighborhood. An animated historical photograph adds generated details that the original photograph never recorded. The output needs to be treated as a generated interpretation, with additional validation for claims about the real world."]},{"title":"Who can use it now","text":["As of October 28, Google describes Genie 4 as a research preview for trusted testers. It says access is coming to Google AI Ultra subscribers through Project Genie and to early access partners using the Gemini API and Google AI Studio. The sources reviewed do not give a verified general release date. A waitlist is available through the official model page. [Access details](https://deepmind.google/models/genie/)","This is not a public model-weight release. Readers can inspect Google’s demonstrations and stated limitations today, but should not assume they can immediately run Genie 4 locally or obtain general API access. Google says generated video and audio will carry its SynthID watermark. That transparency measure does not resolve the accuracy and consistency limits.","The development to watch is whether a short interactive draft becomes dependable enough to help people create, teach or test something useful. Genie 4 expands the kinds of experiences Google can demonstrate. Wider access and independent evaluation will determine how much of that promise carries into everyday work."]}],"limitations":["That moves the experiment closer to an explorable sketch: a place whose appearance, camera movement, atmosphere and conversations can be tried together. Google’s published demonstrations show what the company has selected to present. Bright has not independently tested Genie 4 or measured how consistently it follows instructions.","Google acknowledges that visual details and object permanence can drift, especially during longer sessions, complex edits or navigation. Complex physical interactions, spatial accuracy, legible text, audio accuracy and long-term consistency remain research problems. Its model page says continuous interaction lasts a few minutes, rather than extended sessions. Google’s limitations","As of October 28, Google describes Genie 4 as a research preview for trusted testers. It says access is coming to Google AI Ultra subscribers through Project Genie and to early access partners using the Gemini API and Google AI Studio. The sources reviewed do not give a verified general release date. A waitlist is available through the official model page. Access details","This is not a public model-weight release. Readers can inspect Google’s demonstrations and stated limitations today, but should not assume they can immediately run Genie 4 locally or obtain general API access. Google says generated video and audio will carry its SynthID watermark. That transparency measure does not resolve the accuracy and consistency limits."],"sources":[{"id":"source:google-genie-4-interactive-worlds-audio:1","title":"Google DeepMind · October 27 Genie 4 announcement","type":"institution","url":"https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/"},{"id":"source:google-genie-4-interactive-worlds-audio:2","title":"Google DeepMind · Genie 4 capabilities, access and limitations","type":"institution","url":"https://deepmind.google/models/genie/"}],"links":[]},{"id":"liquid-ai-lfm3-local-multimodal-models","kind":"discovery","path":"/discoveries/liquid-ai-lfm3-local-multimodal-models","title":"Liquid AI releases small models for local text, vision and voice tasks","description":"LFM3 brings downloadable models to everyday devices, with a focus on specialized applications. Hardware requirements, software integration and a custom commercial license still shape what developers can build.","domains":["Open intelligence","Computing","Access"],"published":"2026-10-28","modified":null,"sourceDate":"2026-10-27","dates":{"captureDate":"2026-10-28","eventDate":"2026-10-27","lastReviewedDate":"2026-10-28","publicationDate":"2026-10-27"},"provenance":{"origin":"editorial","externalId":"https://www.liquid.ai/blog/lfm3"},"revisions":[{"id":"liquid-ai-lfm3-local-multimodal-models-initial","recordedAt":"2026-10-28","sourceIds":["source:liquid-ai-lfm3-local-multimodal-models:1","source:liquid-ai-lfm3-local-multimodal-models:2","source:liquid-ai-lfm3-local-multimodal-models:3","source:liquid-ai-lfm3-local-multimodal-models:4"],"summary":"Added source-checked announcement reporting, practical stakes, exact access and licensing caveats, tags and original editorial cover."}],"corrections":[],"review":"Emerging; confidence: unassessed; approved; ai-assisted. Primary announcement and official model/access documentation source-checked October 28. Company claims attributed. No independent hands-on tests. Original AI-generated illustration inspected, explicitly credited and distinguished from model output.","sections":[{"title":"Editorial image","text":["Editorial illustration of a laptop containing a miniature workshop for documents, images and audio, beside a person using a smartphone.: https://brightaifuture.com/media/content/df650f0d2b058c254901a56c9f8a0908b43ec08b700e736d5cb21a30f24b238b.png","Credit: AI-generated editorial illustration for Bright; not model output. License: Original artwork commissioned for Bright.","Source: https://brightaifuture.com/media/content/df650f0d2b058c254901a56c9f8a0908b43ec08b700e736d5cb21a30f24b238b.png","illustration; not an LFM3 screenshot, actual model output, a product performance test or a privacy audit."]},{"title":"Liquid AI releases small models for local text, vision and voice tasks","text":["Liquid AI announced LFM3 on October 27, releasing an open-weight model family built around a small shared backbone. The launch covers text, vision, audio, combined multimodal input and decision models in roughly the two-billion-parameter range. The weights are available through the company’s Hugging Face collection, giving developers something they can download and evaluate now. [Liquid’s announcement](https://www.liquid.ai/blog/lfm3) · [Official collection](https://huggingface.co/collections/LiquidAI/lfm3)","The practical stake is where useful AI runs. A model small enough for a laptop or phone could support tasks such as organizing documents, interpreting an image or responding to a spoken request without sending every input to a remote model. These are building blocks for applications, rather than a complete personal assistant that becomes reliable simply by downloading a file."]},{"title":"Several tools built from a small backbone","text":["The family includes LFM3-1.7B for text, LFM3-VL-2B for vision, LFM3-Audio-2B for sound and speech, and LFM3-Omni-2B for combined vision, audio and language. Liquid also lists d1-2B and d1-Omni-2B decision models, alongside the base model. Developers can select a model for a narrower task instead of running every capability in one larger system. [Release overview](https://www.liquid.ai/blog/lfm3)","The text model card specifies 1.73 billion parameters and a maximum context window of 256,000 tokens across input and output. Parameters are the learned values in a model; fewer of them can reduce the resources needed to run it. Context is the material a model can consider within a request. A large advertised context ceiling does not guarantee that every detail will be recalled correctly or that the longest request will be practical on every device. [Text model card](https://huggingface.co/LiquidAI/LFM3-1.7B)","Liquid says the text model’s weights take about 1 GB at four-bit precision. That is a compressed model-footprint claim, not a promise that the whole application needs only 1 GB of RAM. Runtime software, working memory, context and other components add to the total. The company’s own launch discussion reports additional cache memory for long contexts. Hardware fit needs to be measured with the actual model format, application and workload."]},{"title":"Local processing offers a choice, not automatic privacy","text":["A business could evaluate a model for extracting structured fields from a familiar type of document. A developer could explore voice commands or image questions within an application they control. Liquid supplies local inference options and describes fine-tuning on a single GPU to specialize the models. Bright has not run these models or verified their speed, quality or fine-tuning requirements. [Model formats and guidance](https://huggingface.co/LiquidAI/LFM3-1.7B)","Keeping inference on a device can reduce how much content needs to leave it. Privacy still depends on the surrounding application: search tools, cloud storage, telemetry and external services may transmit information. A local language model connected to an online tool does not make the tool’s activity local. Teams need to check the full data path before promising private or offline operation.","That distinction is especially relevant for agents, software that uses a model to choose and call tools. A model can propose an action, but the application supplies its permissions and executes it. Downloadable weights do not settle whether an agent should be allowed to send messages, change records or handle sensitive files."]},{"title":"Read the benchmark claims with their boundaries","text":["Liquid reports improvements over its earlier models and competitive results against larger models on selected evaluations. Those results are vendor-reported, with different tasks and setups represented in the launch tables. They support a reason to test the model for a particular job; they do not establish that it replaces a larger general-purpose system across workloads. [Published evaluation tables](https://www.liquid.ai/blog/lfm3)","The text model card is more specific about intended use: local agents, tool calling, structured outputs and domain-specific fine-tuning. It recommends retrieval or search for knowledge-heavy work and discourages using the model for agentic coding or complex multi-step reasoning because of its small size. Those limits belong beside the performance claims when deciding whether to adopt it. [Intended use and limitations](https://huggingface.co/LiquidAI/LFM3-1.7B)"]},{"title":"Downloadable does not mean unrestricted","text":["LFM3-1.7B uses the custom LFM Open License v1.0. It is based on Apache 2.0 but adds a commercial-use condition tied to a $10 million annual-revenue threshold. Larger businesses need a separate commercial agreement with Liquid. Redistribution also carries license and attribution requirements, while qualified nonprofits have an exemption for non-commercial or research use. Readers planning commercial deployment should review the full terms for their circumstances. [Liquid’s license](https://www.liquid.ai/lfm-license)","That makes “open-weight” the useful description here. The weights can be obtained, inspected and adapted subject to their license; the release is not simply under standard Apache 2.0 terms. Access to a downloadable model and unrestricted commercial permission are different questions.","For a first evaluation, the useful test is a bounded task with representative examples: does the model produce the required format, handle errors and uncertainty, and meet the device’s memory and response-time needs? LFM3 makes that experiment available now. Its contribution to everyday local AI will be measured by applications that work reliably, not by the small parameter count alone."]}],"limitations":["Liquid says the text model’s weights take about 1 GB at four-bit precision. That is a compressed model-footprint claim, not a promise that the whole application needs only 1 GB of RAM. Runtime software, working memory, context and other components add to the total. The company’s own launch discussion reports additional cache memory for long contexts. Hardware fit needs to be measured with the actual model format, application and workload.","A business could evaluate a model for extracting structured fields from a familiar type of document. A developer could explore voice commands or image questions within an application they control. Liquid supplies local inference options and describes fine-tuning on a single GPU to specialize the models. Bright has not run these models or verified their speed, quality or fine-tuning requirements. Model formats and guidance","Keeping inference on a device can reduce how much content needs to leave it. Privacy still depends on the surrounding application: search tools, cloud storage, telemetry and external services may transmit information. A local language model connected to an online tool does not make the tool’s activity local. Teams need to check the full data path before promising private or offline operation.","The text model card is more specific about intended use: local agents, tool calling, structured outputs and domain-specific fine-tuning. It recommends retrieval or search for knowledge-heavy work and discourages using the model for agentic coding or complex multi-step reasoning because of its small size. Those limits belong beside the performance claims when deciding whether to adopt it. Intended use and limitations","LFM3-1.7B uses the custom LFM Open License v1.0. It is based on Apache 2.0 but adds a commercial-use condition tied to a $10 million annual-revenue threshold. Larger businesses need a separate commercial agreement with Liquid. Redistribution also carries license and attribution requirements, while qualified nonprofits have an exemption for non-commercial or research use. Readers planning commercial deployment should review the full terms for their circumstances. Liquid’s license"],"sources":[{"id":"source:liquid-ai-lfm3-local-multimodal-models:1","title":"Liquid AI · LFM3 release announcement","type":"institution","url":"https://www.liquid.ai/blog/lfm3"},{"id":"source:liquid-ai-lfm3-local-multimodal-models:2","title":"Liquid AI · LFM3-1.7B model card","type":"repository","url":"https://huggingface.co/LiquidAI/LFM3-1.7B"},{"id":"source:liquid-ai-lfm3-local-multimodal-models:3","title":"Liquid AI · Official LFM3 collection","type":"repository","url":"https://huggingface.co/collections/LiquidAI/lfm3"},{"id":"source:liquid-ai-lfm3-local-multimodal-models:4","title":"Liquid AI · LFM Open License v1.0","type":"institution","url":"https://www.liquid.ai/lfm-license"}],"links":[]},{"id":"open-models","kind":"collection","path":"/open-models","title":"Open-Source & Open-Weight AI Models: Compare and Run Locally","description":"Compare 15 AI model families by license, size and openness. Inspect documented local candidates, try a sourced setup guide, and download a dated comparison.","domains":["Open intelligence"],"published":"2026-09-19","modified":"2026-10-11","review":"AI-assisted primary-source review checked 2026-09-19. A curated edition, not a complete live census or independent benchmark reproduction.","limitations":["Model cards and licenses change; every profile names a specific checked release. Hardware calculations describe idealized weight storage, not compatibility. Developer claims remain attributed. Historical examples retain their original source dates."],"sources":[{"id":"osi","title":"OSI · Open Source AI Definition 1.0","url":"https://opensource.org/ai/open-source-ai-definition"},{"id":"cache","title":"Transformers · Cache strategies","url":"https://huggingface.co/docs/transformers/en/kv_cache"},{"id":"quantization","title":"Transformers · Quantization","url":"https://huggingface.co/docs/transformers/en/quantization/bitsandbytes"},{"id":"llamacpp","title":"llama.cpp · local inference and GGUF","url":"https://github.com/ggml-org/llama.cpp"},{"id":"lora","title":"PEFT · Low-rank adaptation","url":"https://huggingface.co/docs/peft/main/en/conceptual_guides/lora"},{"id":"qwen-opening","title":"Qwen3.5-397B-A17B · model card","url":"https://huggingface.co/Qwen/Qwen3.5-397B-A17B"},{"title":"Qwen2.5-7B","url":"https://huggingface.co/Qwen/Qwen2.5-7B"},{"title":"Qwen2.5-7B-Instruct","url":"https://huggingface.co/Qwen/Qwen2.5-7B-Instruct"},{"title":"bartowski / Qwen2.5-7B-Instruct-GGUF","url":"https://huggingface.co/bartowski/Qwen2.5-7B-Instruct-GGUF"},{"title":"DeepSeek-R1-Distill-Qwen-7B","url":"https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B"},{"title":"Qwen2.5-7B-Instruct-AWQ","url":"https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-AWQ"},{"title":"Qwen2.5-7B-Instruct-GPTQ-Int4","url":"https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-GPTQ-Int4"},{"id":"qwen38-repo","title":"Qwen3.8 official repository","url":"https://github.com/QwenLM/Qwen3.8","type":"repository"},{"id":"qwen38-card","title":"Qwen3.8-27B model card","url":"https://huggingface.co/Qwen/Qwen3.8-27B","type":"registry"},{"id":"deepseek-v41-announcement","title":"DeepSeek V4.1 Flash announcement","url":"https://api-docs.deepseek.com/news/news260910/","type":"institution"},{"id":"deepseek-v41-card","title":"DeepSeek-V4.1-Flash model card","url":"https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash","type":"registry"},{"id":"mistral-small4-release","title":"Introducing Mistral Small 4","url":"https://mistral.ai/news/mistral-small-4/","type":"institution"},{"id":"mistral-small4-card","title":"Mistral Small 4 model card","url":"https://huggingface.co/mistralai/Mistral-Small-4-119B-2603","type":"registry"},{"title":"Mistral Small 4 release","url":"https://docs.mistral.ai/models/mistral-small-4-0-26-03"},{"id":"gemma4-release-log","title":"Gemma release log","url":"https://ai.google.dev/gemma/docs/releases","type":"institution"},{"id":"gemma4-release","title":"Gemma 4: Byte for byte, the most capable open models","url":"https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/","type":"institution"},{"id":"gemma4-card","title":"Gemma 4 model card","url":"https://ai.google.dev/gemma/docs/core/model_card_4","type":"report"},{"id":"gptoss-release","title":"Introducing gpt-oss","url":"https://openai.com/index/introducing-gpt-oss/","type":"institution"},{"id":"gptoss-card","title":"gpt-oss model card","url":"https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf","type":"report"},{"id":"llama4-release","title":"The Llama 4 herd","url":"https://ai.meta.com/blog/llama-4-multimodal-intelligence/","type":"institution"},{"id":"olmo3-release","title":"Olmo 3: Charting a path through the model flow","url":"https://allenai.org/blog/olmo3","type":"institution"},{"id":"olmo3-home","title":"Olmo model flow","url":"https://allenai.org/olmo","type":"institution"},{"id":"nemotron3-omni-report","title":"Nemotron 3 Nano Omni technical report","url":"https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Omni-report.pdf","type":"report"},{"id":"nemotron3-omni-page","title":"Nemotron 3 Nano Omni 30B-A3B Reasoning","url":"https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning","type":"institution"},{"id":"nemotron3-family","title":"Nemotron AI models","url":"https://developer.nvidia.com/topics/ai/nemotron","type":"institution"},{"id":"smollm3-blog","title":"SmolLM3: smol, multilingual, long-context reasoner","url":"https://huggingface.co/blog/smollm3","type":"institution"},{"id":"smollm3-card","title":"SmolLM3-3B model card","url":"https://huggingface.co/HuggingFaceTB/SmolLM3-3B","type":"registry"},{"id":"molmo2-media-release","title":"Ai2 media center — Molmo 2 release month","url":"https://allenai.org/media-center","type":"institution"},{"id":"molmo2-repo","title":"Molmo 2 repository","url":"https://github.com/allenai/molmo2","type":"repository"},{"id":"molmo2-readme","title":"Molmo 2 model and training documentation","url":"https://github.com/allenai/molmo2/blob/main/README.md","type":"report"},{"id":"kimi25-release-note","title":"Moonshot AI Kimi K2.5 release note","url":"https://github.com/MoonshotAI/kimi-help-center/blob/master/en-US/agent/swarm.md","type":"repository"},{"id":"kimi25-repo","title":"Kimi K2.5 official repository","url":"https://github.com/MoonshotAI/Kimi-K2.5","type":"repository"},{"id":"kimi25-card","title":"Kimi K2.5 model card","url":"https://huggingface.co/moonshotai/Kimi-K2.5","type":"registry"},{"id":"kimi25-paper","title":"Kimi K2.5: Visual Agentic Intelligence","url":"https://arxiv.org/abs/2602.02276","type":"paper"},{"id":"glm53-release-note","title":"Z.ai GLM-V release note","url":"https://github.com/zai-org/GLM-V","type":"repository"},{"id":"glm53-card","title":"GLM-5.3-Flash model card","url":"https://huggingface.co/zai-org/GLM-5.3-Flash","type":"registry"},{"id":"glm53-vllm","title":"vLLM GLM-5.3-Flash recipe","url":"https://github.com/vllm-project/recipes/blob/main/models/zai-org/GLM-5.3-Flash.yaml","type":"repository"},{"id":"evo2-announcement","title":"Arc Institute Evo 2 announcement","url":"https://arcinstitute.org/news/evo2","type":"institution"},{"id":"evo2-paper","title":"Genome modeling and design across all domains of life with Evo 2","url":"https://arcinstitute.org/manuscripts/Evo2.pdf","type":"paper"},{"id":"evo2-training","title":"Savanna training code linked by Evo 2","url":"https://github.com/zymrael/savanna","type":"repository"},{"id":"evo2-repo","title":"Evo 2 repository","url":"https://github.com/ArcInstitute/evo2","type":"repository"},{"id":"groot-n17-release","title":"Isaac GR00T N1.7 release","url":"https://github.com/NVIDIA/Isaac-GR00T/releases","type":"repository"},{"id":"groot-repo","title":"Isaac GR00T repository","url":"https://github.com/NVIDIA/Isaac-GR00T","type":"repository"},{"id":"groot-platform","title":"Isaac GR00T platform","url":"https://developer.nvidia.com/isaac/gr00t","type":"institution"},{"id":"voxtral-rt-release","title":"Voxtral Transcribe 2 and Realtime release","url":"https://mistral.ai/news/voxtral-transcribe-2/","type":"institution"},{"id":"voxtral-rt-card","title":"Voxtral Mini 4B Realtime model card","url":"https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602","type":"registry"},{"id":"mistral-changelog","title":"Mistral model changelog","url":"https://docs.mistral.ai/resources/changelogs","type":"institution"},{"id":"graphcast-science-2023","title":"Learning skillful medium-range global weather forecasting","url":"https://www.science.org/doi/10.1126/science.adi2336","type":"paper"},{"id":"weathernext-graph-repository","title":"WeatherNext Graph documentation and artifacts","url":"https://github.com/google-deepmind/weathernext/tree/main/docs/weathernext1_graph","type":"repository"},{"id":"ecmwf-aifs-operational-2025","title":"ECMWF’s AI forecasts become operational","url":"https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational","type":"institution"},{"id":"ecmwf-anemoi-models","title":"Anemoi models","url":"https://github.com/ecmwf/anemoi-models","type":"repository"},{"id":"nvidia-earth-2-weather-models-2026","title":"NVIDIA Earth-2 open models for weather and climate","url":"https://blogs.nvidia.com/blog/nvidia-earth-2-open-models/","type":"institution"},{"id":"nvidia-earth-2-research-release-2026","title":"NVIDIA launches Earth-2 family of open models","url":"https://research.nvidia.com/labs/climate/post/2026.01.26-amslaunch-healda-atlas-stormscope/","type":"institution"},{"id":"openvla-repository","title":"OpenVLA: An open-source vision-language-action model","url":"https://github.com/openvla/openvla","type":"repository"},{"id":"openvla-paper","title":"OpenVLA: An Open-Source Vision-Language-Action Model","url":"https://arxiv.org/abs/2406.09246","type":"paper"},{"id":"isaac-groot-n1-7-release","title":"GR00T N1.7 release","url":"https://github.com/NVIDIA/Isaac-GR00T/releases/tag/n1.7-release","type":"repository"},{"id":"esmfold-science-2023","title":"Evolutionary-scale prediction of atomic-level protein structure with a language model","url":"https://www.science.org/doi/10.1126/science.ade2574","type":"paper"},{"id":"esm-repository","title":"Evolutionary Scale Modeling","url":"https://github.com/facebookresearch/esm","type":"repository"},{"id":"esmfold-pubmed-record","title":"Evolutionary-scale prediction of atomic-level protein structure with a language model","url":"https://pubmed.ncbi.nlm.nih.gov/36927031/","type":"government"},{"id":"nvidia-bionemo-organization","title":"NVIDIA BioNeMo repositories","url":"https://github.com/NVIDIA-BioNeMo","type":"repository"},{"id":"bionemo-framework-repository","title":"BioNeMo Framework","url":"https://github.com/NVIDIA/bionemo-framework","type":"repository"},{"id":"bionemo-framework-paper","title":"BioNeMo Framework: a modular, high-performance library for AI model development in drug discovery","url":"https://arxiv.org/abs/2411.10548","type":"paper"},{"id":"mattergen-nature-2025","title":"A generative model for inorganic materials design","url":"https://doi.org/10.1038/s41586-025-08628-5","type":"paper"},{"id":"mattergen-repository","title":"MatterGen code and data release","url":"https://github.com/microsoft/mattergen","type":"repository"},{"id":"prithvi-eo-2-repository","title":"Prithvi-EO-2.0","url":"https://github.com/NASA-IMPACT/Prithvi-EO-2.0","type":"repository"},{"id":"nasa-open-ai-foundation-models","title":"NASA open science AI foundation models","url":"https://www.nas.nasa.gov/SC24/research/project27.php","type":"government"},{"id":"prithvi-eo-2-paper","title":"Prithvi-EO-2.0: A Versatile Multi-Temporal Foundation Model for Earth Observation Applications","url":"https://arxiv.org/abs/2412.02732","type":"paper"},{"id":"segment-anything-repository","title":"Segment Anything","url":"https://github.com/facebookresearch/segment-anything","type":"repository"},{"id":"segment-anything-paper","title":"Segment Anything","url":"https://arxiv.org/abs/2304.02643","type":"paper"},{"id":"medsam-repository","title":"MedSAM: Segment Anything in Medical Images","url":"https://github.com/bowang-lab/MedSAM","type":"repository"},{"id":"medsam-nature-communications","title":"Segment anything in medical images","url":"https://www.nature.com/articles/s41467-024-44824-z","type":"paper"},{"id":"whisper-repository","title":"Whisper","url":"https://github.com/openai/whisper","type":"repository"},{"id":"whisper-release","title":"Introducing Whisper","url":"https://openai.com/index/whisper/","type":"institution"},{"id":"seamless-communication-repository","title":"Seamless Communication","url":"https://github.com/facebookresearch/seamless_communication","type":"repository"},{"id":"seamlessm4t-paper","title":"SeamlessM4T: Massively Multilingual & Multimodal Machine Translation","url":"https://arxiv.org/abs/2308.11596","type":"paper"},{"id":"starcoder2-repository","title":"StarCoder2","url":"https://github.com/bigcode-project/starcoder2","type":"repository"},{"id":"starcoder2-paper","title":"StarCoder 2 and The Stack v2: The Next Generation","url":"https://arxiv.org/abs/2402.19173","type":"paper"},{"id":"gpt-oss-availability","title":"OpenAI open-weight models (gpt-oss)","url":"https://help.openai.com/en/articles/11870455","type":"institution"},{"id":"smolvlm-release","title":"SmolVLM: Redefining small and efficient multimodal models","url":"https://huggingface.co/blog/smolvlm","type":"institution"},{"id":"gemma-3n-model-card","title":"Gemma 3n model card","url":"https://ai.google.dev/gemma/docs/gemma-3n/model_card","type":"institution"},{"id":"gemma-3n-release","title":"Introducing Gemma 3n: The developer guide","url":"https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide/","type":"institution"},{"id":"stable-diffusion-repository","title":"Stable Diffusion","url":"https://github.com/CompVis/stable-diffusion","type":"repository"},{"id":"stable-diffusion-public-release","title":"Stable Diffusion public release","url":"https://stability.ai/news-updates/stable-diffusion-public-release","type":"institution"},{"id":"audiocraft-repository","title":"AudioCraft","url":"https://github.com/facebookresearch/audiocraft","type":"repository"},{"id":"musicgen-model-card","title":"MusicGen model card","url":"https://github.com/facebookresearch/audiocraft/blob/main/model_cards/MUSICGEN_MODEL_CARD.md","type":"repository"},{"id":"musicgen-paper","title":"Simple and Controllable Music Generation","url":"https://arxiv.org/abs/2306.05284","type":"paper"},{"id":"octo-rss-2024-paper","title":"Octo: An Open-Source Generalist Robot Policy","url":"https://octo-models.github.io/paper.pdf","type":"paper"},{"id":"octo-model-repository","title":"Octo generalist robot policy","url":"https://github.com/octo-models/octo","type":"repository"},{"id":"octo-arxiv-paper","title":"Octo: An Open-Source Generalist Robot Policy","url":"https://arxiv.org/abs/2405.12213","type":"paper"},{"id":"foundation-sec-card","title":"Foundation-Sec-8B model card","url":"https://huggingface.co/RedHatAI/Foundation-Sec-8B","type":"registry"},{"id":"foundation-sec-report","title":"Llama-3.1-FoundationAI-SecurityLLM-Base-8B Technical Report","url":"https://arxiv.org/abs/2504.21039","type":"paper"},{"id":"anomalib-paper","title":"Anomalib: A Deep Learning Library for Anomaly Detection","url":"https://arxiv.org/abs/2202.08341","type":"paper"},{"id":"anomalib-repository","title":"Anomalib","url":"https://github.com/open-edge-platform/anomalib","type":"repository"},{"id":"mvtec-ad-dataset","title":"MVTec AD industrial anomaly-detection dataset","url":"https://www.mvtec.com/research-teaching/datasets/mvtec-ad","type":"dataset"},{"id":"pedagogy-open-llm-paper","title":"Narrowing the Gap: Supervised Fine-Tuning of Open-Source LLMs as a Viable Alternative to Proprietary Models for Pedagogical Tools","url":"https://arxiv.org/abs/2507.05305","type":"paper"},{"id":"qwen3-repository","title":"Qwen3","url":"https://github.com/QwenLM/Qwen3","type":"repository"},{"id":"cosmos-3-model-card","title":"Cosmos3-Nano model card","url":"https://build.nvidia.com/nvidia/cosmos3-nano/modelcard","type":"registry"},{"id":"cosmos-3-documentation","title":"Cosmos 3 documentation","url":"https://docs.nvidia.com/cosmos/latest/cosmos3/index.html","type":"institution"},{"id":"cosmos-3-repository","title":"NVIDIA Cosmos repository","url":"https://github.com/NVIDIA/cosmos","type":"repository"},{"title":"GPT-2 made release itself part of the debate","url":"https://openai.com/index/gpt-2-1-5b-release/"},{"title":"A community trained GPT-Neo","url":"https://github.com/EleutherAI/gpt-neo"},{"title":"BLOOM paired scale with visible governance","url":"https://huggingface.co/bigscience/bloom"},{"title":"LLaMA accelerated the open-weight ecosystem","url":"https://arxiv.org/abs/2302.13971"},{"title":"Mistral 7B normalized a compact Apache release","url":"https://github.com/mistralai/mistral-src"},{"title":"OLMo made “fully open” a concrete claim","url":"https://allenai.org/olmo2"},{"title":"DeepSeek-R1 released reasoning weights","url":"https://github.com/deepseek-ai/DeepSeek-R1"},{"title":"Hugging Face Hub","url":"https://huggingface.co/docs/hub/index"},{"title":"MLX","url":"https://github.com/ml-explore/mlx"},{"title":"Ollama","url":"https://github.com/ollama/ollama"},{"title":"vLLM","url":"https://github.com/vllm-project/vllm"},{"title":"SGLang","url":"https://github.com/sgl-project/sglang"},{"title":"Transformers","url":"https://github.com/huggingface/transformers"},{"title":"PEFT","url":"https://github.com/huggingface/peft"},{"title":"LM Evaluation Harness","url":"https://github.com/EleutherAI/lm-evaluation-harness"},{"title":"Open WebUI","url":"https://github.com/open-webui/open-webui"}],"sections":[{"title":"Download the comparison","text":["Bright Open Models comparison, edition 2026-10-11-r2, https://brightaifuture.com/open-models/comparison/2026-10-11-r2.json","Not established means this review did not verify it, not that it is absent.","CSV: https://brightaifuture.com/open-models/comparison/2026-10-11-r2.csv","JSON: https://brightaifuture.com/open-models/comparison/2026-10-11-r2.json","The export publication date is separate from each record source check. File size is not runtime memory or proof of compatibility."]},{"title":"From weights to possibility","text":["Model weights are the learned numbers that can be downloaded, compressed, run and modified. The opening uses a hypothetical 8B-total, 2B-active teaching model: idealized 16/8/4-bit weight storage is 16/8/4 decimal GB. It is not a released checkpoint. Active work is different from total storage; cache, runtime and packaging need extra memory. A sourced historical example remains available: Qwen3.5-397B-A17B reports 397B total language-model parameters and 17B active, excluding its vision encoder."]},{"title":"Weights","text":["The learned numbers a model uses to transform inputs into outputs. Downloading them makes running a compatible model yourself possible; it does not tell you everything about how it was trained.","Source: https://opensource.org/ai/open-source-ai-definition"]},{"title":"Open-weight AI","text":["A useful descriptive label for released model parameters. It says little on its own about rights, training code or data. Read the license and the artifact list separately.","Source: https://opensource.org/ai/open-source-ai-definition"]},{"title":"Open source AI","text":["OSI’s definition asks for freedoms to use, study, modify and share, plus parameters, code and sufficiently detailed data information. It does not require every training datum to be redistributable. Bright shows the components rather than treating a download as certification.","Source: https://opensource.org/ai/open-source-ai-definition"]},{"title":"Source-available","text":["Source can be visible while its terms restrict use or redistribution. Availability and permission answer different questions. “Open” is used inconsistently across the industry; the actual artifacts and terms are the useful comparison.","Source: https://opensource.org/ai/open-source-ai-definition"]},{"title":"Permissive / restrictive licenses","text":["Permissive terms generally allow broad reuse with conditions such as preserving notices. Other releases add use, territory or scale restrictions. Code and weights can have different licenses; a family name is not a license.","Source: https://opensource.org/ai/open-source-ai-definition"]},{"title":"Architecture / inference code","text":["Architecture describes how the model is assembled. Inference code runs it. Having both helps others implement a runtime, but does not supply the process that produced the weights.","Source: https://opensource.org/ai/open-source-ai-definition"]},{"title":"Training code / recipe / data","text":["Training code executes learning; a recipe records choices such as data processing and settings. Data information explains what went in. Access to the actual data is a separate question, including its rights and any gaps.","Source: https://opensource.org/ai/open-source-ai-definition"]},{"title":"Evaluation","text":["A useful result names the exact model, task, scoring method and runtime conditions. A developer’s table is evidence of what that developer reports, not independent reproduction. This edition does not publish a benchmark leaderboard.","Source: https://huggingface.co/Qwen/Qwen3.5-397B-A17B"]},{"title":"Quantization","text":["Store numbers using fewer bits. This reduces weight storage, but accuracy and speed depend on the method, model and backend. A theoretical 4-bit estimate is not the byte size of a packaged download.","Source: https://huggingface.co/docs/transformers/en/quantization/bitsandbytes"]},{"title":"GGUF","text":["A model-file format used by llama.cpp and compatible tools. It can contain different tensor types and quantizations. “GGUF” by itself does not tell you the precision, license, quality or required memory.","Source: https://github.com/ggml-org/llama.cpp"]},{"title":"LoRA / adapters","text":["Train a comparatively small set of additional parameters while keeping the base weights fixed. An adapter needs its compatible base model; it is not usually a standalone replacement for the full model.","Source: https://huggingface.co/docs/peft/main/en/conceptual_guides/lora"]},{"title":"RAM / VRAM / KV cache","text":["System RAM and dedicated GPU memory are different budgets; unified memory is shared. Attention caches retain information about the current context. Longer conversations and more simultaneous users can consume much more memory. CPU offloading trades GPU memory for system memory and transfer work.","Source: https://huggingface.co/docs/transformers/en/kv_cache"]},{"title":"Atlas edition","text":["Qwen3.8: Qwen3.8-27B. A downloadable 27B vision-language member of the Qwen3.8 generation, selected here because its Apache-2.0 terms are materially simpler than the custom license on the largest Qwen3.8 checkpoint. Apache-2.0. Vendor benchmark claims need independent reproduction Licensing differs across Qwen3.8 sizes","DeepSeek V4.1 Flash: DeepSeek-V4.1-Flash. A downloadable DeepSeek multimodal Flash model, with sparse activation that differs between prompt processing and generation. MIT. Displayed file size and stated backbone parameter count differ Training corpus is not released","Mistral Small 4: Mistral-Small-4-119B-2603. A hybrid sparse model that combines instruction following, reasoning, visual understanding, and coding modes in one downloadable checkpoint. Apache-2.0. Training data is not documented in enough detail for reconstruction Performance claims are primarily vendor-reported","Gemma 4: Gemma 4 26B-A4B. Google’s Gemma 4 generation spans edge and workstation sizes; this atlas record uses the 26B mixture-of-experts checkpoint as its reference point. Apache-2.0. Audio is limited to the E2B, E4B, and 12B variants Training data is described but not released","gpt-oss: gpt-oss-120b and gpt-oss-20b. Two open-weight text reasoning models with adjustable reasoning effort, structured output, and tool-use support. Apache-2.0, with a separate usage policy. Text-only Training corpus is not released Self-hosters own system-level safeguards","Llama 4: Scout and Maverick. Meta’s first natively multimodal Llama generation and first Llama mixture-of-experts release. Llama 4 Community License Agreement. Custom license is not OSI-approved Training data and full training stack are not released","Olmo 3: Olmo 3 7B and 32B. A fully open model flow spanning base, mid-trained, long-context, instruction, reasoning, and RL-zero checkpoints. Apache-2.0 for models and code; data components retain their stated licenses. Text-only Full reproduction still requires substantial compute","Nemotron 3 Nano Omni: Nemotron-3-Nano-Omni-30B-A3B-Reasoning. A compact sparse reasoning model that accepts video, audio, images, and text while activating about three billion parameters per token. NVIDIA Open Model Agreement. Custom license Complete multimodal training corpus is not released Vendor benchmarks need independent reproduction","SmolLM3: SmolLM3-3B. A compact multilingual reasoning model whose engineering blueprint, mixtures, training frameworks, and alignment process are published. Apache-2.0. Text-only Six explicitly supported languages Smaller capacity than frontier-scale models","Molmo 2: Molmo2-4B, Molmo2-8B, and Molmo2-O-7B. A family of open vision-language models for images, multiple images, and video, with explicit spatial outputs such as pointing and object tracking. Apache-2.0 repository and model terms; some source datasets require separate agreements. Some training datasets need separate agreements Spatial outputs still require task-specific evaluation before deployment","Kimi K2.5: Kimi-K2.5. A trillion-parameter visual agentic model with a comparatively small active footprint and public weights and code. Modified MIT License. Extremely large total checkpoint Custom license condition at large commercial scale Training data is not released","GLM-5.3 Flash: GLM-5.3-Flash. A natively multimodal sparse GLM checkpoint released under MIT terms for long-context agent and coding work. MIT. Training data is not disclosed in enough detail Large server-class checkpoint","Evo 2: Evo 2 7B and 40B. A biological foundation model trained across genomes for sequence understanding and design, released with code, weights, and OpenGenome2 data. Apache-2.0 for code and model artifacts; dataset terms must be checked separately. Research model, not a clinical diagnostic Generated biological sequences require domain safety review","Isaac GR00T: GR00T N1.7. An open vision-language-action model for generalized humanoid skills, accompanied by reference code, data tooling, and a robotics development stack. NVIDIA Open Model License / repository-specific terms. Robot results depend on hardware, calibration, and safety controls Real-world operation needs physical safeguards","Voxtral Realtime: Voxtral-Mini-4B-Realtime-2602. A compact streaming speech-recognition model designed for low-latency transcription through a realtime API. Apache-2.0. Speech-focused rather than a general multimodal assistant Training data is not fully disclosed"]},{"title":"Ways into the atlas","text":["Explore every family. All 15 checked family records, without a ranking.","Run AI on my own machine. Families with smaller named checkpoints. Memory is only one constraint: explore the hardware planner before choosing a runtime.","Build a coding or reasoning system. Publishers describe coding or reasoning capabilities. Evaluate your own tasks; these are starting points, not winners.","Work with images, audio or video. Different input modalities mean different tools. Open a profile to see exactly which ones a checkpoint accepts.","Build robotics or physical AI. Start with a documented vision-language-action family, then follow the Machines world into research examples.","Explore biology and scientific research. Evo models biological sequences. The Life and Earth worlds below contain further specialist applications.","Inspect training and reproduce research. These records document substantial training artifacts. Inspect each door: released components and permissions still differ.","Understand model licenses. Compare exact release terms across all families. “Downloadable” does not mean every use is permitted.","Explore small models. These families include checkpoints of eight billion parameters or fewer. Small does not mean universally compatible.","Explore very large models. These records describe checkpoints above 100 billion total parameters or a larger documented backbone. Sparse activation does not remove storage requirements."]},{"title":"Six application worlds","text":["Earth: Read a changing planet. Atmospheres · weather · observation","Life: Look inside living systems. Proteins · biology · medical images · materials","Machines: Give a learned pattern a physical task. Robotics · simulation · inspection","Voice: Carry words across a boundary. Speech · translation · audio","Creation: See an image as something editable. Images · vision · multimodal tools","Work: Bring a model into a discipline. Code · reasoning · security · education"]},{"title":"What can I run?","text":["The device-first planner starts with a MacBook, Windows laptop, desktop/gaming PC, NVIDIA GPU, workstation/server or phone/edge device. Visitors choose memory, 4/8/16-bit precision and an overhead reserve. Advanced controls explain context and CPU offload without pretending to compute exact cache or combine RAM and VRAM. Decimal GB = total parameter billions × bits / 8. A weight budget is not a compatibility, speed or quality guarantee."]},{"title":"Documented descendants","text":["Base model: Qwen2.5-7B. The shared starting checkpoint. This small historical branch makes the relationship inspectable. https://huggingface.co/Qwen/Qwen2.5-7B","Instruction adaptation: Qwen2.5-7B-Instruct. Qwen’s instruction-tuned checkpoint builds on the base model. https://huggingface.co/Qwen/Qwen2.5-7B-Instruct","Community quantization: bartowski / Qwen2.5-7B-Instruct-GGUF. A community maintainer packages that instruction model in GGUF quantizations for compatible runtimes. https://huggingface.co/bartowski/Qwen2.5-7B-Instruct-GGUF","Reasoning distillation: DeepSeek-R1-Distill-Qwen-7B. DeepSeek documents fine-tuning Qwen2.5-Math-7B on R1-generated samples. This is a separate Qwen Math branch, not a descendant of the generic Qwen2.5-7B checkpoint shown above. https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B","Official AWQ quantization: Qwen2.5-7B-Instruct-AWQ. Qwen publishes an AWQ quantization of its instruction-tuned checkpoint. It changes the numerical representation, not the declared base lineage. https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-AWQ","Official GPTQ Int4 quantization: Qwen2.5-7B-Instruct-GPTQ-Int4. The publisher identifies this as a GPTQ 4-bit quantization of Qwen2.5-7B-Instruct. Its model tree explicitly names the base and instruction checkpoints. https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-GPTQ-Int4","Separately, Cisco describes Foundation-Sec-8B as continued pretraining from Llama-3.1-8B on cybersecurity text: https://huggingface.co/fdtn-ai/Foundation-Sec-8B. These are distinct branches. Shared architecture or family names alone do not establish a weight lineage."]},{"title":"Why people care","text":["Keep it close. A clinic, a newsroom, a person. Prompts can stay on controlled hardware when the runtime, logs, plugins and network behavior are configured that way. Local weights alone do not guarantee privacy.","Keep it working. A place without a reliable connection. After downloading the necessary artifacts, supported local inference can work offline. Tools that fetch information or call services still need a network.","Make it specific. A language, a discipline, a task. Fine-tunes and adapters let builders change behavior. They need suitable data, rights, evaluation and a license that permits the intended use.","Keep a choice. An institution choosing its infrastructure. Self-hosting can give companies, researchers and countries operational control and portability. Hardware, expertise and supply chains remain dependencies.","Look inside. A researcher asking why it failed. Available weights and code enable forms of inspection and experimentation that an API alone cannot. Reproducibility also needs the missing recipe, data and evaluation conditions."]},{"title":"Real-world examples","text":["Forecasting the atmosphere ten days ahead: GraphCast turns a recent state of the atmosphere into a global medium-range forecast. Its implementation and pretrained weights can be downloaded, so researchers can inspect and run a forecasting model outside a hosted product. Performance figures come from the peer-reviewed paper by the model's developers. The repository publishes Apache-2.0 code and CC-BY-4.0 listed weights; it does not publish a complete training-data recreation. Limit: This is a research forecast driven by prepared atmospheric inputs, not an operational warning service. The public release includes code and weights, but not a complete disclosed training stack.","An AI forecast enters daily operations: ECMWF moved its Artificial Intelligence Forecasting System into its operational forecast suite in February 2025 and publishes model code and checkpoints through the Anemoi project. Operational status and performance claims are ECMWF's account of its own service; the separately inspectable Anemoi repository establishes public implementation artifacts, not independent validation of the operating service. Limit: The operating organization reports its own performance. Repository, checkpoint, and input-data terms vary by release, so the full service should not be described as one uniformly open-source system.","Weather models that can be carried into local workflows: NVIDIA’s January 2026 Earth-2 announcement separated atmospheric initialization, global forecasting and local nowcasting. Atlas and StormScope were available then; HealDA was announced for later release. Availability and capability descriptions are NVIDIA's release claims. “Open weights” describes downloadable parameters under NVIDIA terms, not a complete open training stack. Limit: This is a developer-described release, not evidence that the components improve a public forecast service. The January source distinguishes released models from a planned release; later availability requires a separate check. Individual artifact terms control reuse.","Teaching a robot from demonstrations: OpenVLA gives robotics teams a pretrained vision-language-action model plus instructions for fine-tuning it on their own robot demonstrations. The repository documents a reproducible research artifact. Its MIT code license does not extend to the released model checkpoints. Limit: Repository code is MIT, while checkpoints inherit Llama 2 Community License restrictions. Laboratory evaluations do not establish safe unattended operation.","A shared starting point for humanoid skills: Isaac GR00T N1.7 is a downloadable vision-language-action model intended as a base for developing humanoid robot skills. Capabilities are described by NVIDIA in its own repository. Public code and downloadable weights are separate artifacts with separate terms. Limit: The weights use the NVIDIA Open Model License, which is distinct from an OSI-approved open-source license. Public artifacts do not establish site-specific reliability or physical safety.","Predicting a protein’s shape from its sequence: ESMFold applies a protein language model to infer three-dimensional structure directly from an amino-acid sequence, with checkpoints and bulk prediction tools available to researchers. The peer-reviewed paper and public artifacts establish the research result; neither establishes experimental validity for every predicted protein. Limit: Predicted structures vary in confidence and require scientific follow-up. The MIT repository and CC-BY-4.0 Atlas do not amount to a complete released training corpus and recipe.","Assembling biological model workflows: BioNeMo brings biological foundation models and training components into one framework for protein, molecular, and drug-research experiments. This record describes public developer infrastructure. It makes no claim that a BioNeMo workflow has produced an effective therapy. Limit: Availability and licenses vary by component, and model weights generally use NVIDIA Open Model License terms. The repositories establish tooling, not a successful drug or clinical outcome.","Proposing materials for a desired property: MatterGen generates candidate inorganic crystal structures while conditioning on properties a researcher wants, changing the starting point from searching a known catalog to proposing structures for testing. The result is peer-reviewed and accompanied by public code. Restricted source data means the released artifacts are not a complete reproducible training stack. Limit: Generated candidates still require synthesis and physical measurement. Repository code is MIT, but some ICSD-derived training material cannot be redistributed, preventing complete recreation from the public package; checkpoint terms still require their own model-card review.","Adapting one Earth model to many maps: Prithvi-EO-2.0 is a geospatial foundation model that can be adapted to Earth-observation tasks such as flood mapping, crop classification, and land-use analysis. NASA and IBM document the public artifacts and research tasks. This broader mapping record is distinct from Bright's separate Prithvi in-orbit demonstration. Limit: The Prithvi-EO-2.0 repository code is MIT and the named 300M checkpoint is Apache-2.0, but those terms do not make every satellite input freely redistributable. Performance can shift by region, sensor, season, and local ground truth.","SAM 1: selecting an object with a point or a box: The original 2023 Segment Anything release, now commonly called SAM 1, turns a click, box, or rough mask into an object segmentation, giving annotators and analysts an editable first pass across varied images. This is the historical SAM 1 release, not a claim about later Segment Anything generations. The official repository establishes Apache-2.0 model artifacts and intended workflows; SA-1B data access and downstream image rights remain separate. Limit: A plausible mask may still be wrong, especially in unfamiliar imagery. Apache-2.0 model artifacts do not settle the rights attached to input images, dataset access, or downstream uses.","Giving medical-image annotators an editable first boundary: MedSAM adapts promptable segmentation to medical images and provides interactive tools for delineating structures or lesions for research annotation workflows. The public research repository establishes an inspectable workflow. No clinical deployment or medical outcome is inferred. Limit: The MedSAM repository is Apache-2.0; that license is not evidence of diagnostic accuracy, regulatory clearance, or safe clinical use, and medical-image dataset rights remain separate.","Transcribing speech on infrastructure you control: Whisper provides downloadable speech-recognition models for transcription, language identification, translation into English, and caption-making without requiring a hosted speech service. The official repository establishes MIT-licensed code and weights. That unusually permissive artifact release remains distinct from access to the training data. Limit: Accuracy varies with language, accent, noise, recording conditions, and subject matter. The training corpus and a complete training recipe were not released.","Moving between speech and text across languages: SeamlessM4T joins speech recognition, text translation, speech translation, and speech generation in one multilingual research model family. Capabilities and language coverage are developer-described. MIT code and noncommercial model weights have materially different reuse permissions. Limit: SeamlessM4T v1 and v2 weights are CC-BY-NC-4.0, so commercial use cannot be assumed. Translation quality, toxicity, and speech identity require evaluation for each language and setting.","Adapting a code model to a private codebase: StarCoder2 offers several downloadable code-model sizes plus fine-tuning and evaluation examples for local completion and generation experiments. The project repository establishes model and tooling availability; benchmark results do not establish correctness in a production codebase. Limit: Generated code can be insecure, incorrect, or incompatible with a project. Repository code is Apache-2.0, while StarCoder2 model weights use BigCode OpenRAIL-M v1; dataset provenance and terms remain separate.","Running a reasoning model on your own machine: OpenAI's gpt-oss family provides downloadable reasoning-model weights for self-hosted tool use, structured outputs, and adjustable reasoning effort. Availability, hardware guidance, and benchmark comparisons are OpenAI's descriptions of its own release. Bright labels this open weight rather than a complete Open Source AI training stack. Limit: Apache-2.0 weights and public inference references do not include the full training data or recipe. Benchmark and memory figures are vendor results tied to the stated quantization and workload.","Asking questions about an image on a small device: SmolVLM is a compact vision-language model family designed for document, image, and visual-question tasks where memory and compute are limited. Artifact availability and efficiency claims come from the publisher's release materials; they do not establish reliability in a particular accessibility workflow. Limit: A small footprint does not guarantee factual answers, accessibility, or adequate speed on every device. Derived checkpoints must be checked separately.","Combining image, audio, video, and text at the edge: Gemma 3n is a multimodal model designed to accept text, images, video, and audio while producing text on resource-constrained devices. Intended capabilities and evaluations are reported by Google. Public weights under Gemma terms are described here as open weight, not as a fully open training stack. Limit: Weights are governed by Gemma terms, not an OSI-approved software license asserted here; training data is summarized rather than released. Device fit and quality vary by hardware, language, and task.","Adapting an image model to a visual practice: Stable Diffusion's downloadable latent-diffusion weights and adaptation code allowed artists and developers to build local image generation, inpainting, fine-tuning, and LoRA workflows. The repository establishes the original public artifacts. Broad ecosystem use does not erase the model license's restrictions or settle training-data disputes. Limit: CreativeML OpenRAIL-M includes use-based restrictions and is not a simple permissive software license. Training-data rights, consent, bias, and the rights of generated work require project-specific review.","Sketching music from words and melody: MusicGen generates music from a text description and can use an input melody as an additional condition, providing a downloadable tool for research and creative prototyping. The official code and model card document a research tool. MIT code and noncommercial weights are deliberately reported as different layers. Limit: Code is MIT, but model weights are CC-BY-NC-4.0 and therefore do not authorize commercial product use. The model card cautions against downstream deployment without risk evaluation.","Adapting one robot policy across nine platforms: Octo is a generalist manipulation policy trained on the Open X-Embodiment dataset and evaluated as a reusable starting point for robots with different sensors, action spaces, and physical forms. The reported cross-platform results come from the model authors and were published at RSS 2024. The project exposes MIT-licensed code and downloadable Octo 1.5 checkpoints; those artifacts do not establish unattended robot safety. Limit: The experiments are research evaluations on specific tasks and hardware, not evidence of safe autonomous deployment. Repository code and the named Octo 1.5 checkpoint are MIT; the Open X-Embodiment source datasets still require their own provenance and terms review.","A security model for a SOC's own evidence: Foundation-Sec-8B is a cybersecurity-focused Llama 3.1 derivative that organizations can download and adapt for security operations work involving their own alerts, cases, and threat knowledge. The use cases and evaluation are Cisco Foundation AI's account, not an independently corroborated deployment. The Foundation-Sec-8B card lists Apache-2.0 for this checkpoint; its Llama lineage and every downstream artifact still require version-specific terms review. Limit: The operational workflows and benchmark results come from the model publisher, with no independent evidence that it improves outcomes in a live security operations center. Static training data and adversarial inputs also make current threat intelligence and guarded deployment essential.","A local visual-inspection workbench: Anomalib packages multiple anomaly-detection methods into a modular library for finding and localizing unusual regions in inspection images, with paths from experiments to optimized edge inference. The paper and maintained repository establish an open implementation and deployment toolkit. They do not establish that every bundled or compatible model weight is open, or that a factory has validated the resulting inspection system. Limit: Anomalib is an Apache-2.0 software toolkit, not one universally licensed set of open weights. Component models, pretrained assets, and datasets retain their own terms, and benchmark behavior does not establish performance on a particular production line.","Explaining a novice programmer's compiler error: Researchers fine-tuned compact downloadable language models to explain C compiler errors using examples derived from real introductory-programming mistakes. The evaluation and pedagogical comparison are reported by the study authors. “Open-source” is their terminology; Bright separately records that the evaluated Qwen and Llama variants carry different model licenses and are not a single uniform open stack. Limit: This is an author-reported research evaluation, not a school deployment or evidence of improved learning outcomes. Dataset access and the distinct Qwen and Llama model terms must be checked before reuse.","Simulating a scene before a robot enters it: Cosmos 3 provides downloadable omnimodal world-model variants that can process or generate combinations of text, images, video, audio, and action sequences for physical-AI research. Capabilities and limitations come from NVIDIA documentation and the publisher model card. The code repository is Apache-2.0, while Cosmos 3 model artifacts use OpenMDW 1.1; Bright does not collapse those separate terms into one “open source” claim. Limit: The publisher warns that outputs can violate physical laws, lose object state, drift over long horizons, and fail on safety-critical edge cases. Generated scenes are not ground truth or evidence that a robot policy is safe in the physical world."]},{"title":"Open and hosted tradeoffs","text":["Sensitive records. Controlled local infrastructure can keep inputs in one place.. A hosted service may have suitable contractual and technical controls.. Inspect logs, network calls, access rules and retention in either setup.","A small team shipping a feature. Weights allow modification and a choice of runtimes.. An API can remove much of the hardware and serving work.. Compare the whole workload: latency, staffing, reliability and cost.","Specialist research. Weights and training artifacts allow deeper experiments.. A hosted frontier model may perform better on the chosen task.. Evaluate with representative data; availability is not capability.","Long-term operation. A retained checkpoint can help reproduce a prior system.. Managed updates and safety services can reduce operational work.. You still need security, monitoring, evaluation and incident response."]},{"title":"Timeline","text":["2019-02: GPT-2 made release itself part of the debate. OpenAI released GPT-2 in stages and published the full 1.5B model in November 2019. The weights became public; the release was not a complete open training stack.","2021-03: A community trained GPT-Neo. EleutherAI released a community-built language-model family with public code and checkpoints, creating a durable reference for work outside the largest labs.","2022-07: BLOOM paired scale with visible governance. BigScience released a multilingual model built by an international collaboration with unusually extensive data and governance documentation; its RAIL terms differ from a permissive license.","2022-08: Stable Diffusion made image weights adaptable. The public 1.x weights catalyzed local image tools, fine-tunes, and LoRAs while bringing data rights and use-restricted licensing into public view.","2022-09: Whisper put speech code and weights under MIT. The release made local transcription and translation practical for many builders, while accuracy and speech-data questions remained setting-specific.","2023-02: LLaMA accelerated the open-weight ecosystem. Meta published the first LLaMA research models under a research-oriented license, showing that downloadable weights can still carry substantial restrictions.","2023-09: Mistral 7B normalized a compact Apache release. A small, commercially usable model helped expand local serving and adaptation; the exact version and license still matter for every descendant.","2024-02: OLMo made “fully open” a concrete claim. Ai2 released training data, code, checkpoints, recipe, and evaluation artifacts, giving researchers more than final weights to inspect and reproduce.","2024-10: OSI defined Open Source AI 1.0. The definition separated freedoms and required components into a testable vocabulary. It is a standard for analysis, not a legal judgment on every model.","2025-01-20: DeepSeek-R1 released reasoning weights. DeepSeek released R1, R1-Zero, and six distilled checkpoints, renewing questions about distillation, reproducibility, and post-release safety. Its own code and R1 weights are MIT; Llama-derived distills retain their base-model terms.","2025-08-05: OpenAI released the gpt-oss weights. The 120b and 20b reasoning models arrived with Apache-2.0 weights and inference references, but without a complete public training-data corpus or recipe.","2026-03-31: Gemma 4 moved the family to Apache 2.0. Google’s release log records the initial Gemma 4 weights under Apache 2.0 on March 31; Google published its announcement article on April 2. The release expanded commercial reuse rights for the weights; it did not publish the complete training data or training recipe."]},{"title":"People and infrastructure","text":["Hugging Face Hub: Hosts model files, datasets, cards, metadata, and revision history. A listing exposes provenance clues; it does not prove a derivative is valid or inherits the same permissions.","llama.cpp: Runs supported model formats locally across consumer hardware and exposes quantization choices that determine memory, speed, and output fidelity.","MLX: Provides an array and machine-learning framework optimized for Apple silicon, supporting local training and inference experiments.","Ollama: Packages and serves supported models on a local machine; convenience does not change the model license or validate its answers.","vLLM: Serves language models with continuous batching and memory-management techniques for higher-throughput inference.","SGLang: Supplies a serving engine and programming interface for structured model workloads; serving behavior remains separate from model permissions and quality.","Transformers: Provides common model definitions, loading, training, and inference APIs across many families, making revisions easier to run and compare.","PEFT: Creates lightweight adapters such as LoRAs so a base model can be specialized without retraining every parameter; each adapter still needs base-model and data provenance.","LM Evaluation Harness: Runs language-model evaluations with recorded tasks and settings. A comparable score still requires the exact model revision, prompt, precision, harness version, and date.","Open WebUI: Adds a browser interface around local or remote model servers. The interface is its own product with privacy, extension, and security behavior beyond the model."]}],"links":[{"title":"Documentation-led Mac and Windows setup guide","path":"/guides/run-ai-on-your-computer"},{"title":"CSV comparison","path":"/open-models/comparison.csv"},{"title":"JSON comparison","path":"/open-models/comparison.json"},{"title":"Qwen3.8: license, size and what is open","path":"/open-models/qwen"},{"title":"DeepSeek V4.1 Flash: license, size and what is open","path":"/open-models/deepseek"},{"title":"Mistral Small 4: license, size and what is open","path":"/open-models/mistral"},{"title":"Gemma 4: license, size and what is open","path":"/open-models/gemma"},{"title":"gpt-oss: license, size and what is open","path":"/open-models/gpt-oss"},{"title":"Llama 4: license, size and what is open","path":"/open-models/llama"},{"title":"Olmo 3: license, size and what is open","path":"/open-models/olmo"},{"title":"Nemotron 3 Nano Omni: license, size and what is open","path":"/open-models/nemotron"},{"title":"SmolLM3: license, size and what is open","path":"/open-models/smollm"},{"title":"Molmo 2: license, size and what is open","path":"/open-models/molmo"},{"title":"Kimi K2.5: license, size and what is open","path":"/open-models/kimi"},{"title":"GLM-5.3 Flash: license, size and what is open","path":"/open-models/glm"},{"title":"Evo 2: license, size and what is open","path":"/open-models/evo"},{"title":"Isaac GR00T: license, size and what is open","path":"/open-models/groot"},{"title":"Voxtral Realtime: license, size and what is open","path":"/open-models/voxtral"},{"title":"Forecasting the atmosphere ten days ahead","path":"/discoveries/graphcast"},{"title":"An AI forecast enters daily operations","path":"/discoveries/aifs"},{"title":"Weather models that can be carried into local workflows","path":"/discoveries/nvidia-earth-2"},{"title":"Teaching a robot from demonstrations","path":"/discoveries/openvla-robot-manipulation"},{"title":"A shared starting point for humanoid skills","path":"/discoveries/groot-humanoid-skills"},{"title":"Predicting a protein’s shape from its sequence","path":"/discoveries/esmfold-protein-structures"},{"title":"Assembling biological model workflows","path":"/discoveries/bionemo-biological-workflows"},{"title":"Proposing materials for a desired property","path":"/discoveries/mattergen-material-candidates"},{"title":"Adapting one Earth model to many maps","path":"/discoveries/prithvi-earth-observation"},{"title":"SAM 1: selecting an object with a point or a box","path":"/discoveries/segment-anything-annotation"},{"title":"Giving medical-image annotators an editable first boundary","path":"/discoveries/medsam-medical-segmentation"},{"title":"Transcribing speech on infrastructure you control","path":"/discoveries/whisper-local-transcription"},{"title":"Moving between speech and text across languages","path":"/discoveries/seamlessm4t-speech-translation"},{"title":"Adapting a code model to a private codebase","path":"/discoveries/starcoder2-local-code"},{"title":"Running a reasoning model on your own machine","path":"/discoveries/gpt-oss"},{"title":"Asking questions about an image on a small device","path":"/discoveries/smolvlm-on-device-vision"},{"title":"Combining image, audio, video, and text at the edge","path":"/discoveries/gemma-3n-edge-assistance"},{"title":"Adapting an image model to a visual practice","path":"/discoveries/stable-diffusion-adapted-creation"},{"title":"Sketching music from words and melody","path":"/discoveries/musicgen-conditioned-music"},{"title":"Adapting one robot policy across nine platforms","path":"/discoveries/octo-robot-policy-adaptation"},{"title":"A security model for a SOC's own evidence","path":"/discoveries/foundation-sec-8b"},{"title":"A local visual-inspection workbench","path":"/discoveries/anomalib-industrial-inspection"},{"title":"Explaining a novice programmer's compiler error","path":"/discoveries/open-model-pedagogy-compiler-errors"},{"title":"Simulating a scene before a robot enters it","path":"/discoveries/cosmos-3-physical-ai-prototyping"},{"title":"Open intelligence immersive experience","path":"/open-intelligence"},{"title":"Open thread","path":"/threads/open"}]}],"nextOffset":null}