Qwen3.5 4B is the shared writing starting option because its exact local package and task documentation support one short exercise across these reviewed profiles. A larger download is not a reason to promote another model.
Candidates, not a ranking. Known documented downloads come first, smallest first; unknown sizes follow. Different checkpoints and formats have different requirements.
Hugging Face · Commercial use permitted under Apache-2.0; review accompanying terms.
1.92 GB listed GGUF file. Selected ggml-org conversion of Hugging Face SmolLM3, not a maker-published artifact. This 1.92 GB listing is rounded, not total runtime memory or a census of all quantizations.
Follow the documentation-led LM Studio guide, then check the fictional picnic reply against your notes.
Qwen / Alibaba Cloud · Commercial use permitted under Apache-2.0; review accompanying terms.
About 9.32 GB of original safetensors shards, plus supporting files. A smaller named family member, distinct from the Qwen3.8-27B grid checkpoint. Not runtime memory.
The publisher documents Transformers, vLLM and SGLang. Check exact architecture and image support in your backend.
OpenAI · Commercial use permitted under Apache-2.0; review accompanying terms.
About 13.76 GB of listed model shards. Smaller member named in this family record. Sum of the three rounded shard listings, excluding supporting files. Other artifact formats differ; this is not runtime memory.
Check Harmony-format support in the documented runtime, and verify every generated fact against your supplied material.
Running a reasoning model on your own machine → · Emerging. 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.
Qwen / Alibaba Cloud · Commercial use permitted under Apache-2.0; review accompanying terms.
About 9.32 GB of original safetensors shards, plus supporting files. A smaller named family member, distinct from the Qwen3.8-27B grid checkpoint. Not runtime memory.
The publisher documents Transformers, vLLM and SGLang. Check exact architecture and image support in your backend.
Allen Institute for AI (Ai2) · Commercial use permitted under Apache-2.0; review accompanying terms.
About 19.40 GB of selected documented weights. Molmo2-4B only, not 8B or O-7B. Four float32 shards include vision components, so 4B label is not whole tensor parameter count. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.
Read the exact checkpoint and runtime documentation before downloading. Bright has not tested speed, runtime memory or output quality on your computer.
Asking questions about an image on a small device → · Emerging. A small footprint does not guarantee factual answers, accessibility, or adequate speed on every device. Derived checkpoints must be checked separately.
Mistral AI · Commercial use permitted under Apache-2.0; review accompanying terms.
About 8.86 GB of selected documented weights. 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.
Read the exact checkpoint and runtime documentation before downloading. Bright has not tested speed, runtime memory or output quality on your computer.
Google DeepMind · Commercial use permitted under Apache-2.0; review accompanying terms.
16 GB listed original safetensors file, plus supporting files. Smaller family member, distinct from the 26B-A4B grid checkpoint. Not runtime memory.
The official card documents audio input and transcription examples. Confirm audio processor and runtime support; this is distinct from the grid’s image/text-only 26B checkpoint.
Transcribing speech on infrastructure you control → · Demonstrated. Accuracy varies with language, accent, noise, recording conditions, and subject matter. The training corpus and a complete training recipe were not released.
Qwen / Alibaba Cloud · Commercial use permitted under Apache-2.0; review accompanying terms.
About 9.32 GB of original safetensors shards, plus supporting files. A smaller named family member, distinct from the Qwen3.8-27B grid checkpoint. Not runtime memory.
The publisher documents Transformers, vLLM and SGLang. Check exact architecture and image support in your backend.
OpenAI · Commercial use permitted under Apache-2.0; review accompanying terms.
About 13.76 GB of listed model shards. Smaller member named in this family record. Sum of the three rounded shard listings, excluding supporting files. Other artifact formats differ; this is not runtime memory.
Check Harmony-format support in the documented runtime, and verify every generated fact against your supplied material.
Mistral AI · Commercial use permitted under Apache-2.0; review accompanying terms.
About 70.81 GB of selected documented weights. Official NVFP4 variant, all 13 consolidated shards. Excludes separate FP8 repository and alternate representations. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.
Read the exact checkpoint and runtime documentation before downloading. Bright has not tested speed, runtime memory or output quality on your computer.
Adapting a code model to a private codebase → · Emerging. 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.
Google DeepMind · Commercial use permitted under Apache-2.0; review accompanying terms.
1.49 GB listed original safetensors file, plus supporting files. Specialist Gemma embedding model, distinct from the grid checkpoint. Index storage and runtime memory are separate.
For builders: produces search vectors, not chat replies. Read query/document-prefix guidance and use supported bfloat16 or float32, not float16. A search application is still needed.
OpenAI · Commercial use permitted under Apache-2.0; review accompanying terms.
About 13.76 GB of listed model shards. Smaller member named in this family record. Sum of the three rounded shard listings, excluding supporting files. Other artifact formats differ; this is not runtime memory.
A text-generation candidate for answers from excerpts supplied by a separate retrieval application. It does not create a search index or retrieve your files by itself.
A security model for a SOC's own evidence → · Emerging. 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.
OpenAI · Commercial use permitted under Apache-2.0; review accompanying terms.
About 13.76 GB of listed model shards. Smaller member named in this family record. Sum of the three rounded shard listings, excluding supporting files. Other artifact formats differ; this is not runtime memory.
Check Harmony-format support in the documented runtime, and verify every generated fact against your supplied material.
Allen Institute for AI (Ai2) · Commercial use permitted under Apache-2.0; review accompanying terms.
About 14.60 GB of selected documented weights. 7B Instruct only, not 32B, Think or Olmo 3.1. Three shards. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.
A transparent model flow for research builders. Supply your own source excerpts and verify citations; weights do not provide a literature-search service.
Explaining a novice programmer's compiler error → · Demonstrated. 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.
NVIDIA · Custom or artifact-specific terms: read them. Review the model and data licenses for the intended robot, dataset, and deployment.
About 6.91 GB of selected documented weights. General checkpoint only, excludes task fine-tunes, scheduler.pt and training_args.bin. Preserve both card and bundled license references. Robotics specialist, not ordinary laptop chat recommendation. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.
For robotics builders: inspect the model card and bundled NVIDIA license, robot interface, calibration and physical safeguards. This is not an ordinary laptop chat recommendation; Bright has not run the checkpoint.
Arc Institute and collaborators · Commercial use permitted under Apache-2.0; review accompanying terms.
About 13.77 GB of selected documented weights. evo2_7b.pt long-context 7B only, not separate base or 40B; not loaded or executed. Specialist sequence model, not ordinary laptop chat recommendation. Selected weight files only; supporting files and runtime memory are separate. Not a smallest-download census or proof of hardware compatibility.
For genomics researchers: read the Evo 2 sequence-model documentation and specialist hardware requirements. This is not an ordinary laptop chat recommendation; Bright has not run the checkpoint.
A shared starting point for humanoid skills → · Experimental. 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.
How we know · documentation and why people care
Original six task selections, sources checked 2026-10-10. No models or hardware tested by Bright.
HuggingFaceTB/SmolLM3-3B
A small text candidate for the fictional picnic exercise above.
The publisher documents llama.cpp and other local backends. The ggml-org GGUF page names LM Studio and llama.cpp for its conversion.
SmolLM3-Q4_K_M.gguf: listed at 1.92 GB. This is one converted file, not the original checkpoint or total runtime memory.
Apache 2.0. Check the conversion and app version; generated facts still need review.
Compare the exact GGUF file with your app’s supported formats, then try the fictional notes.
An instruction-tuned multimodal candidate; a model’s modalities do not automatically appear in every app.
Google’s card demonstrates Transformers with its multimodal processor. Check the selected runtime’s image and audio support separately.
Original model.safetensors: listed at 16 GB, plus supporting files. “E4B” is an effective count: the card reports 4.5B effective and 8B including embeddings, not a download size.
Apache 2.0 on the checked card. Verify current access terms and the exact artifact; total runtime memory and speed are unknown.
Choose one input type and confirm the runtime supports it before downloading a conversion.
A text reasoning candidate for people ready for a larger download.
The publisher documents Ollama, LM Studio and Transformers. The runtime must handle the required Harmony format.
Root safetensors shards total about 13.76 GB (4.79 + 4.80 + 4.17), roughly 14 GB, with MXFP4 expert weights. Other artifact formats differ.
Apache 2.0 plus the linked usage policy. A vendor’s 16 GB memory claim is not a guarantee on a busy 16 GB computer; cache, software and other apps also need memory.
Inspect your app’s exact artifact and memory guidance. Keep connected search and other tools separate from local text inference.
For builders: produces vectors for retrieval and similarity, rather than chat replies. A search application is still needed.
The card demonstrates SentenceTransformers and Transformers. Use bfloat16 where supported or float32; do not use FP16 (float16), which can produce invalid or degraded vectors.
Original model.safetensors: listed at 1.49 GB, plus supporting files. Index storage and runtime memory are separate.
Apache 2.0. Search quality, precision support and handling of your collection need their own checks.
Read the numerical-precision and query/document-prefix guidance before building a local search workflow.
For research builders: turns a question and retrieved scientific excerpts into a cited report. It does not supply a literature-search app.
The card shows Transformers/vLLM code and a prescribed prompt. A beginner-ready desktop runtime has not been verified.
Four safetensors shards: about 16.38 GB (4.90 + 4.92 + 4.98 + 1.58), plus supporting files. The repository also includes duplicate-format .bin weights; you do not need both formats.
Apache 2.0; intended for research/education under Ai2’s responsible-use guidance. Based on Qwen3-8B, with inherited base limitations. Verify citations against the supplied excerpts.
Read Ai2’s workflow and the required prompt before considering this specialist checkpoint.
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.
Look up your chip, installed memory and operating system manually. The planner below uses numbers you enter; it does not inspect your device. If you do not know your specifications, keep them unknown and follow its help links.
Unified memory is shared by CPU and GPU. Dedicated VRAM is separate from system RAM. The memory available after the operating system and other apps is smaller than the installed capacity. File size is a separate disk-space question.
Start with a documented candidate, then check the exact file and your app.
Where do I find my memory?
Mac: Apple menu → System Settings → General → About. On Apple silicon the CPU and GPU share installed memory. Activity Monitor shows current memory use. Apple system information ↗ · Apple memory guide ↗
Windows: Settings → System → About → Device specifications shows installed RAM. Task Manager → Performance → GPU distinguishes dedicated from shared GPU memory. Shared GPU memory uses RAM; it is not extra dedicated VRAM. Leave unclear specifications unknown. Microsoft device guide ↗ · Microsoft GPU explanation ↗
Documented candidates, not a compatibility result
Installed memory is unknown. No memory comparison is made.
These are download sizes, not runtime memory requirements. Supporting files add to the download; the app and context cache need memory beyond weights. Installed capacity is not memory available now; RAM and dedicated VRAM remain separate. Compatibility, runtime memory, speed and output quality are unknown. Bright has not tested these candidates on its own hardware.
Showing 3 of 6 documented candidates, ordered by file size.
How we know: file evidence and the existing API arithmetic
Documented files above come from the existing source-checked task records. Formats are distinct artifacts. A parameter-count estimate is not a download, and neither one measures runtime memory. Compatibility requires checking the exact checkpoint, architecture, app version and backend. Shorter context can reduce cache use, but no cache estimate is inferred here.
The existing model API serves a separate 4-bit parameter-storage screen with a 20% planning reserve. It does not inspect these downloads. This disclosure calls that same API implementation, so its candidates and numbers agree for identical inputs. The reserve is a planning assumption, not a measurement of free memory. For a discrete GPU it uses VRAM alone.
No API arithmetic while the relevant memory pool is unknown. For a NVIDIA desktop enter dedicated VRAM separately.
Buying a computer for this? Inspect the documented configurations.
Documentation checked 2026-10-10. These examples are not purchase recommendations or Bright-tested hardware.
These examples explain tradeoffs, not what to buy. Choose your task and software first; arithmetic alone cannot promise an exact fit or acceptable speed.
Portable unified memory / MacBook Pro
Apple lists M5, M5 Pro and M5 Max variants. M5 offers 16/24/32 GB; selected M5 Max configurations with a 40-core GPU offer up to 128 GB unified memory.
A screen and battery travel with you. CPU, GPU, macOS and apps share that installed memory; a higher capacity does not establish model speed or battery endurance under inference.
Match the exact chip, GPU configuration and installed memory to your runtime’s requirements.
Apple lists M6 and M5 Pro. M6 starts at 16 or 24 GB and supports up to 32 GB; M5 Pro starts at 24 GB and offers 48 or 64 GB unified memory.
A stationary computer with shared CPU/GPU memory; plan separately for a display and peripherals. An exact chip/configuration matters more than the product name.
Check the selected configuration and software support; installed memory is not all available to a model.
GB10 Grace Blackwell with 128 GB coherent unified system memory. NVIDIA announced a partner-only 64 GB configuration for October 23, 2026; that is a future availability date at this check.
An Arm developer system with NVIDIA’s software stack, rather than a portable everyday computer. Vendor “up to” model-capacity claims do not prove speed, fit for every format or Bright-tested performance.
Check the exact 64/128 GB configuration, availability and Arm/backend compatibility for your workflow.
NVIDIA’s desktop RTX 5070 has 12 GB GDDR7; the desktop RTX 5070 Ti has 16 GB. Laptop GPUs with similar names are different products.
Dedicated GPU VRAM and system RAM are separate budgets. Offloading can use system RAM with transfer and speed tradeoffs; card capacity alone does not establish a model fit.
Read the exact desktop card’s VRAM and the PC’s system RAM, power/cooling and runtime support separately.
Eight billion total parameters, two billion active: a hypothetical teaching model, not a released checkpoint. Each of 64 blocks represents 125 million learned values. Volume follows weight storage, not architecture. Active parameters reduce work per token; inactive experts still need storage.
Decimal GB = total parameter billions × bits ÷ 8: 16 GB at FP16/BF16, 8 GB at 8-bit, 4 GB at 4-bit. Cache, runtime, activations and packaging need more. Quantization changes numerical representation, not parameter count.
A real historical example: Qwen3.5-397B-A17B reports 397B language-model parameters and 17B active. Its idealized weight storage is 794 / 397 / 198.5 GB at 16 / 8 / 4 bits, excluding the vision encoder. These are developer-reported counts, not a comparison of capability.
Can it keep the facts straight?
A fictional exercise, not a Bright-tested benchmark. No model runs on this page.
Check the reply yourself
Fluent wording can still hide a wrong fact. Keep the notes beside the reply.
Compare with an illustrative message
The picnic has moved to Sunday at 1 p.m. at River Park. Lee is bringing fruit and Sam is bringing sandwiches. The rain plan is still undecided.
Illustrative wording written for this exercise; not model output. Other wording can preserve the same facts.
Local is a processing choice.
Local inference runs the model on your device. Connected cloud models, web search and other tools can send inputs elsewhere. LM Studio’s app privacy policy ↗ distinguishes these modes. Check the selected model and tools before entering personal information.
Saved chats remain files on your computer; synced folders and backups can create copies elsewhere. Check your storage and backup settings too.
Optional: after downloading a model and its runtime, disconnect and try the fictional prompt. The offline documentation ↗ describes what works without connectivity. This tryout is not a privacy audit and proves nothing about data handling after reconnecting.
This privacy explanation was checked on October 9, 2026. Policies can change; it is separate from the comparison edition.
Source checks remain dated September 19, 2026. The October 11 export edition does not recheck all nine findings. Recheck flags reflect age at the fixed export edition date, not a live clock or a new review. A check more than 45 days before that edition is marked due for recheck.
Downloadable learned parameters for local inspection, adaptation, or inference.
Architecture
A public description or configuration detailed enough to understand the model design.
Inference code
Code or documented integrations for running the released weights.
Training code
Code used to pre-train or post-train the model, beyond inference and fine-tuning examples.
Training recipe
The stages, hyperparameters, mixtures, and other instructions needed to reproduce training.
Data information
A meaningful account of training-data composition, collection, filtering, and governance.
Training data
The actual training corpus, or a reproducible path to reconstruct it.
Evaluation
Published evaluation results, prompts, harnesses, or code that can be examined and repeated.
Commercial use
License terms that state whether and under what conditions commercial deployment is allowed.
What this edition can and cannot tell you.
Sources checked 2026-09-19. Profiles distinguish code, weights, data and permissions. An unknown field means the review did not establish it. It does not mean the artifact does not exist. Model cards are mutable; a checked date is not a promise of continuing availability.
Most release claims come from their developers. Peer review is identified where present, but is not the same as independent reproduction. This guide publishes no cross-model performance ranking. Deeper independent reporting is still needed on multilingual behavior, real-device energy and latency, accessibility outcomes, long-term maintenance and commercial deployment costs.
NVIDIA appears where its released work supports an example. The broader NVIDIA collection remains a separate editorial connection. This destination’s examples and model records span many organizations.
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 ↗
Open-weight AI
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 ↗
Open source AI
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 ↗
Source-available
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 ↗
Permissive / restrictive licenses
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 ↗
Architecture / inference code
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 ↗
Training code / recipe / data
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 ↗
Evaluation
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 ↗
Quantization
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 ↗
GGUF
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 ↗
LoRA / adapters
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 ↗
RAM / VRAM / KV cache
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 ↗
04 / The model families
Many starting points.
Each card names a family. Its record names the checked release, sources and limits. Family membership is not a shared license or performance claim.
33B listed deployment parameters on NVIDIA Build; the same page’s architecture text says 31B A3B, while the model identifier names a 30B-A3B backbone. Accounting differs across first-party fields.
NVIDIA Open Model Agreement · Check use conditions
The bar shows total parameter count on a logarithmic scale, not capability. Unknown sizes have no bar. Compare up to three checked releases.
05 / What people build
Beyond the chat box.
24 documented examples. Research artifacts, operational systems and possible uses retain their separate evidence states and dates.
01 / 4 documented examplesEarthRead a changing planet.Atmospheres · weather · observationEnter this world ＋
Climate · Demonstrated
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.
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.
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.
The peer-reviewed Science paper reported that GraphCast produced skillful ten-day forecasts and outperformed ECMWF's deterministic HRES system across most of the paper's evaluated targets.
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.
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.
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.
ECMWF reports that AIFS became operational on 25 February 2025, producing forecasts four times each day for public and member-state use.
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.
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.
Availability and capability descriptions are NVIDIA's release claims. “Open weights” describes downloadable parameters under NVIDIA terms, not a complete open training stack.
NVIDIA’s January announcement made Medium Range (Atlas) and Nowcasting (StormScope) available and said Global Data Assimilation (HealDA) was expected later that year. It did not establish all three as downloadable on the announcement date.
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.
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.
NASA and IBM document the public artifacts and research tasks. This broader mapping record is distinct from Bright's separate Prithvi in-orbit demonstration.
NASA and IBM publish model variants, code, examples, and task evaluations through the Prithvi repository and model cards.
02 / 4 documented examplesLifeLook inside living systems.Proteins · biology · medical images · materialsEnter this world ＋
Biology · Demonstrated
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.
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.
The peer-reviewed paper and public artifacts establish the research result; neither establishes experimental validity for every predicted protein.
The Science study used the approach to help create the ESM Metagenomic Atlas, a public collection of predicted structures for hundreds of millions of metagenomic proteins.
BioNeMo brings biological foundation models and training components into one framework for protein, molecular, and drug-research experiments.
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.
This record describes public developer infrastructure. It makes no claim that a BioNeMo workflow has produced an effective therapy.
NVIDIA-BioNeMo publishes framework code, example workflows, and links to model weights that developers can combine in research pipelines.
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 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.
The public research repository establishes an inspectable workflow. No clinical deployment or medical outcome is inferred.
The university-led repository publishes a checkpoint, command-line inference, notebooks, and a graphical interface for research use.
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.
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.
The result is peer-reviewed and accompanied by public code. Restricted source data means the released artifacts are not a complete reproducible training stack.
The Nature paper reports generated candidates and an experimental synthesis case; Microsoft released implementation and data-processing material for research use.
03 / 5 documented examplesMachinesGive a learned pattern a physical task.Robotics · simulation · inspectionEnter this world ＋
Robotics · Experimental
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.
Repository code is MIT, while checkpoints inherit Llama 2 Community License restrictions. Laboratory evaluations do not establish safe unattended operation.
The repository documents a reproducible research artifact. Its MIT code license does not extend to the released model checkpoints.
The official release provides checkpoints, LoRA and full fine-tuning paths, configuration files, and evaluation instructions for supported robot environments.
Isaac GR00T N1.7 is a downloadable vision-language-action model intended as a base for developing humanoid robot skills.
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.
Capabilities are described by NVIDIA in its own repository. Public code and downloadable weights are separate artifacts with separate terms.
NVIDIA publishes the repository, inference and fine-tuning tooling, and model weights for research and developer experimentation.
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 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.
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.
The Robotics: Science and Systems 2024 paper reports fine-tuning experiments across nine robot platforms, and the project publishes pretrained Octo 1.5 checkpoints plus training, fine-tuning, inference, and real-robot evaluation code.
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.
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.
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.
The Anomalib paper and repository provide training, evaluation, visualization, and OpenVINO optimization tools that let teams build and test a local anomaly-detection pipeline.
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.
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.
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.
NVIDIA publishes Cosmos 3 model artifacts and local-inference examples, including a DROID robot-policy variant. The reviewed repository marks post-training recipes and task-specific evaluation as coming soon; publication of weights does not establish those planned artifacts as released.
04 / 3 documented examplesVoiceCarry words across a boundary.Speech · translation · audioEnter this world ＋
Access · Demonstrated
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.
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.
OpenAI released model weights, inference code, and a command-line interface under MIT terms in September 2022, making local runs broadly reproducible.
SeamlessM4T joins speech recognition, text translation, speech translation, and speech generation in one multilingual research model family.
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.
Meta publishes inference tooling, checkpoints, and supported-language documentation that allow researchers to run and evaluate the model family.
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.
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.
The official code and model card document a research tool. MIT code and noncommercial weights are deliberately reported as different layers.
Meta's AudioCraft repository publishes inference code, model cards, and MusicGen checkpoints for experimentation.
05 / 4 documented examplesCreationSee an image as something editable.Images · vision · multimodal toolsEnter this world ＋
Infrastructure · Demonstrated
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.
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.
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.
Meta released SAM 1 inference code, Apache-2.0 model checkpoints, notebooks, and separate access to the SA-1B dataset in April 2023.
SmolVLM is a compact vision-language model family designed for document, image, and visual-question tasks where memory and compute are limited.
A small footprint does not guarantee factual answers, accessibility, or adequate speed on every device. Derived checkpoints must be checked separately.
Artifact availability and efficiency claims come from the publisher's release materials; they do not establish reliability in a particular accessibility workflow.
Hugging Face publishes checkpoints, demonstrations, training recipes, tools, and supporting VLM datasets under Apache-2.0 terms for the described release.
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.
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.
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.
Google publishes weights and a detailed model card describing supported inputs, intended uses, evaluations, and deployment considerations.
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.
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.
The repository establishes the original public artifacts. Broad ecosystem use does not erase the model license's restrictions or settle training-data disputes.
The 2022 public release supplied model weights and code that became a base for a large ecosystem of interfaces and adaptations.
06 / 4 documented examplesWorkBring a model into a discipline.Code · reasoning · security · educationEnter this world ＋
Infrastructure · Emerging
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.
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.
The project repository establishes model and tooling availability; benchmark results do not establish correctness in a production codebase.
BigCode publishes 3B, 7B, and 15B model releases, a fine-tuning example, and links to a code evaluation harness.
OpenAI's gpt-oss family provides downloadable reasoning-model weights for self-hosted tool use, structured outputs, and adjustable reasoning effort.
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.
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.
OpenAI released gpt-oss-120b and gpt-oss-20b weights, reference inference code, and tool examples in August 2025.
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 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.
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.
The technical report evaluates the released model on cybersecurity benchmarks, and its public model card documents intended security-operations workflows and downloadable weights.
Researchers fine-tuned compact downloadable language models to explain C compiler errors using examples derived from real introductory-programming mistakes.
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.
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.
The research evaluation combined expert review with automated analysis of 8,000 responses and reported that fine-tuning improved the pedagogical quality of the smaller models to levels comparable with much larger models.
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.
Fine-tuned on R1-generated samplesDeepSeek-R1-Distill-Qwen-7BInspect
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.
R1 supplies training outputs. It is the teacher, not this checkpoint’s base architecture.
LLAMA / A DIFFERENT FAMILY
Separate starting checkpointLlama-3.1-8B
Continued pretraining · cybersecurityFoundation-Sec-8BInspect
Cisco describes continued pretraining on cybersecurity text, producing a domain-adapted base model. This is not merely a smaller file or a prompt wrapper.
Qwen2.5-7B → Qwen2.5-7B-Instruct: instruction adaptation. That instruction checkpoint → bartowski GGUF: community quantization. Source links and the publisher’s stated relationships:
Qwen2.5-7B — The shared starting checkpoint. This small historical branch makes the relationship inspectable.
Qwen2.5-7B-Instruct — Qwen’s instruction-tuned checkpoint builds on the base model.
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.
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.
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.
Separately, Foundation-Sec-8B’s card names Llama-3.1-8B as its base and continued pretraining as the transformation. No edge joins the Llama and Qwen roots.
Fine-tuning changes trained weights; continued pretraining extends the training objective on more data. Distillation learns from a teacher. Quantization changes numerical representation. LoRA trains adapters; it is not itself a descendant we can plot without a named artifact. An architecture derivative or a later family release is not automatically a weight descendant. Applications combine models and software; they need their own evidence.
This is a small documented sample, not a census, endorsement, or count of all descendants. Each artifact retains its own version and license conditions.
NVIDIA appears where its released work supports an example. The broader NVIDIA collection remains a separate editorial connection. These records span many organizations.
06 / What is new
A record worth returning to.
Dated developments from Bright’s evidence ledger. Publication, release and source-check dates mean different things; this is a curated list.
Reflection AI’s Beam preview could expand open-weight AI. What its planned release and reported NVIDIA talks could mean for developers and researchers.
Emerging · The announcement describes a preview undergoing final red-teaming and evaluations.
Downloadable models can handle some everyday tasks on a device you own. That gives people a new choice about their prompts. Artists’ consent, land-use decisions and electric bills raise different questions.
Emerging · Source-based reporting; no hands-on test performed.
A US test is planned for later October. Here are the practical questions for users.
Experimental · Readers should keep those questions separate: who receives information, what the platform uses internally and which choices are available. Bright’s conversation-privacy explainer takes the same practical approach to understanding where a conversation goes and who can access it.
Emerging · This analysis draws on the Council’s published agenda and exhibits, clearly labeled written submissions, and attributed reporting on the hearing. Reported testimony is identified as such.
Germany’s Aleph Alpha released a downloadable German-English model on October 3. It offers another route to self-hosted AI, while showing why open weights, open-source development and easy local use are different things.
Emerging · Open-weight: the model-repository Apache 2.0 grant covers weights and configuration files, not a complete training stack or corpus. The separate inference plugin has its own license.
A chatbot can make it easy to think out loud. The harder question is who else can access those thoughts. For people exploring personal questions, working through an unfinished idea or handling confidential material, where an AI runs is a practical privacy choice.
Emerging · The Florida account remains attributed reporting. A charge does not establish guilt. Bright has not independently reviewed the arrest report, and the reporting contains no case-specific Anthropic response.
AI agents can write code, use tools and keep working long after a person steps away. NVIDIA’s September 28 Open Agent Safety Platform launch puts permission controls outside the agent’s workload and offers a separate hardware-backed watchdog. The aim is to give people stronger control over what agents can reach; the launch does not establish that agents are now safe in every setting.
Emerging · This is an Emerging security-platform record based on NVIDIA’s announcement, technical explanation and documentation. The cited launch materials do not establish independently measured reductions in incidents across real-world deployments.
Access changed. Then what people could build changed.
Selected historical milestones, retaining the different release choices.
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.
EleutherAI released a community-built language-model family with public code and checkpoints, creating a durable reference for work outside the largest labs.
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.
Meta published the first LLaMA research models under a research-oriented license, showing that downloadable weights can still carry substantial restrictions.
The definition separated freedoms and required components into a testable vocabulary. It is a standard for analysis, not a legal judgment on every model.
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.
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.
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.
Researchers publish artifacts. Maintainers make them run. Dataset builders, evaluators and independent quantizers make the work reusable. Hardware and hosting still matter.
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.
Supplies a serving engine and programming interface for structured model workloads; serving behavior remains separate from model permissions and quality.
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.
Runs language-model evaluations with recorded tasks and settings. A comparable score still requires the exact model revision, prompt, precision, harness version, and date.
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.
BRIGHT WEEKLY
Five source-checked stories of AI progress, plus one useful thing to try.
A concise weekly read about progress in everyday life, work and science. We check the sources and make clear what’s demonstrated, newly announced or still being tested.
