# Open Models — intelligence you can build on

A living guide to open-weight AI: explore model families, inspect licenses and training artifacts, understand local memory, and trace what people build from released weights.

Canonical: https://brightaifuture.com/open-models
Format: collection
Source publication: Not established
Bright publication: 2026-09-19
Substantive update: None recorded
Evidence and review: AI-assisted primary-source review checked 2026-09-19. A curated edition, not a complete live census or independent benchmark reproduction.

## From weights to possibility

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.

## Weights

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

## 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: https://opensource.org/ai/open-source-ai-definition

## 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: https://opensource.org/ai/open-source-ai-definition

## 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: https://opensource.org/ai/open-source-ai-definition

## 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: https://opensource.org/ai/open-source-ai-definition

## 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: https://opensource.org/ai/open-source-ai-definition

## 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: https://opensource.org/ai/open-source-ai-definition

## 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: https://huggingface.co/Qwen/Qwen3.5-397B-A17B

## 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: https://huggingface.co/docs/transformers/en/quantization/bitsandbytes

## 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: https://github.com/ggml-org/llama.cpp

## 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: https://huggingface.co/docs/peft/main/en/conceptual_guides/lora

## 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: https://huggingface.co/docs/transformers/en/kv_cache

## Atlas edition

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

## Ways into the atlas

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.

## Six application worlds

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

## What can I run?

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.

## Documented descendants

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.

## Why people care

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.

## Real-world examples

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.

## Open and hosted tradeoffs

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.

## Timeline

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.

## People and infrastructure

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.

## 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.

## Provenance and history

{}

## Original sources

- [OSI · Open Source AI Definition 1.0](https://opensource.org/ai/open-source-ai-definition)
- [Transformers · Cache strategies](https://huggingface.co/docs/transformers/en/kv_cache)
- [Transformers · Quantization](https://huggingface.co/docs/transformers/en/quantization/bitsandbytes)
- [llama.cpp · local inference and GGUF](https://github.com/ggml-org/llama.cpp)
- [PEFT · Low-rank adaptation](https://huggingface.co/docs/peft/main/en/conceptual_guides/lora)
- [Qwen3.5-397B-A17B · model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)
- [Qwen2.5-7B](https://huggingface.co/Qwen/Qwen2.5-7B)
- [Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct)
- [bartowski / Qwen2.5-7B-Instruct-GGUF](https://huggingface.co/bartowski/Qwen2.5-7B-Instruct-GGUF)
- [DeepSeek-R1-Distill-Qwen-7B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B)
- [Qwen2.5-7B-Instruct-AWQ](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-AWQ)
- [Qwen2.5-7B-Instruct-GPTQ-Int4](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-GPTQ-Int4)
- [Qwen3.8 official repository](https://github.com/QwenLM/Qwen3.8)
- [Qwen3.8-27B model card](https://huggingface.co/Qwen/Qwen3.8-27B)
- [DeepSeek V4.1 Flash announcement](https://api-docs.deepseek.com/news/news260910/)
- [DeepSeek-V4.1-Flash model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash)
- [Introducing Mistral Small 4](https://mistral.ai/news/mistral-small-4/)
- [Mistral Small 4 model card](https://huggingface.co/mistralai/Mistral-Small-4-119B-2603)
- [Mistral Small 4 release](https://docs.mistral.ai/models/mistral-small-4-0-26-03)
- [Gemma release log](https://ai.google.dev/gemma/docs/releases)
- [Gemma 4: Byte for byte, the most capable open models](https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/)
- [Gemma 4 model card](https://ai.google.dev/gemma/docs/core/model_card_4)
- [Introducing gpt-oss](https://openai.com/index/introducing-gpt-oss/)
- [gpt-oss model card](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)
- [The Llama 4 herd](https://ai.meta.com/blog/llama-4-multimodal-intelligence/)
- [Olmo 3: Charting a path through the model flow](https://allenai.org/blog/olmo3)
- [Olmo model flow](https://allenai.org/olmo)
- [Nemotron 3 Nano Omni technical report](https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Omni-report.pdf)
- [Nemotron 3 Nano Omni 30B-A3B Reasoning](https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning)
- [Nemotron AI models](https://developer.nvidia.com/topics/ai/nemotron)
- [SmolLM3: smol, multilingual, long-context reasoner](https://huggingface.co/blog/smollm3)
- [SmolLM3-3B model card](https://huggingface.co/HuggingFaceTB/SmolLM3-3B)
- [Ai2 media center — Molmo 2 release month](https://allenai.org/media-center)
- [Molmo 2 repository](https://github.com/allenai/molmo2)
- [Molmo 2 model and training documentation](https://github.com/allenai/molmo2/blob/main/README.md)
- [Moonshot AI Kimi K2.5 release note](https://github.com/MoonshotAI/kimi-help-center/blob/master/en-US/agent/swarm.md)
- [Kimi K2.5 official repository](https://github.com/MoonshotAI/Kimi-K2.5)
- [Kimi K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5)
- [Kimi K2.5: Visual Agentic Intelligence](https://arxiv.org/abs/2602.02276)
- [Z.ai GLM-V release note](https://github.com/zai-org/GLM-V)
- [GLM-5.3-Flash model card](https://huggingface.co/zai-org/GLM-5.3-Flash)
- [vLLM GLM-5.3-Flash recipe](https://github.com/vllm-project/recipes/blob/main/models/zai-org/GLM-5.3-Flash.yaml)
- [Arc Institute Evo 2 announcement](https://arcinstitute.org/news/evo2)
- [Genome modeling and design across all domains of life with Evo 2](https://arcinstitute.org/manuscripts/Evo2.pdf)
- [Savanna training code linked by Evo 2](https://github.com/zymrael/savanna)
- [Evo 2 repository](https://github.com/ArcInstitute/evo2)
- [Isaac GR00T N1.7 release](https://github.com/NVIDIA/Isaac-GR00T/releases)
- [Isaac GR00T repository](https://github.com/NVIDIA/Isaac-GR00T)
- [Isaac GR00T platform](https://developer.nvidia.com/isaac/gr00t)
- [Voxtral Transcribe 2 and Realtime release](https://mistral.ai/news/voxtral-transcribe-2/)
- [Voxtral Mini 4B Realtime model card](https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602)
- [Mistral model changelog](https://docs.mistral.ai/resources/changelogs)
- [Learning skillful medium-range global weather forecasting](https://www.science.org/doi/10.1126/science.adi2336)
- [WeatherNext Graph documentation and artifacts](https://github.com/google-deepmind/weathernext/tree/main/docs/weathernext1_graph)
- [ECMWF’s AI forecasts become operational](https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational)
- [Anemoi models](https://github.com/ecmwf/anemoi-models)
- [NVIDIA Earth-2 open models for weather and climate](https://blogs.nvidia.com/blog/nvidia-earth-2-open-models/)
- [NVIDIA launches Earth-2 family of open models](https://research.nvidia.com/labs/climate/post/2026.01.26-amslaunch-healda-atlas-stormscope/)
- [OpenVLA: An open-source vision-language-action model](https://github.com/openvla/openvla)
- [OpenVLA: An Open-Source Vision-Language-Action Model](https://arxiv.org/abs/2406.09246)
- [GR00T N1.7 release](https://github.com/NVIDIA/Isaac-GR00T/releases/tag/n1.7-release)
- [Evolutionary-scale prediction of atomic-level protein structure with a language model](https://www.science.org/doi/10.1126/science.ade2574)
- [Evolutionary Scale Modeling](https://github.com/facebookresearch/esm)
- [Evolutionary-scale prediction of atomic-level protein structure with a language model](https://pubmed.ncbi.nlm.nih.gov/36927031/)
- [NVIDIA BioNeMo repositories](https://github.com/NVIDIA-BioNeMo)
- [BioNeMo Framework](https://github.com/NVIDIA/bionemo-framework)
- [BioNeMo Framework: a modular, high-performance library for AI model development in drug discovery](https://arxiv.org/abs/2411.10548)
- [A generative model for inorganic materials design](https://doi.org/10.1038/s41586-025-08628-5)
- [MatterGen code and data release](https://github.com/microsoft/mattergen)
- [Prithvi-EO-2.0](https://github.com/NASA-IMPACT/Prithvi-EO-2.0)
- [NASA open science AI foundation models](https://www.nas.nasa.gov/SC24/research/project27.php)
- [Prithvi-EO-2.0: A Versatile Multi-Temporal Foundation Model for Earth Observation Applications](https://arxiv.org/abs/2412.02732)
- [Segment Anything](https://github.com/facebookresearch/segment-anything)
- [Segment Anything](https://arxiv.org/abs/2304.02643)
- [MedSAM: Segment Anything in Medical Images](https://github.com/bowang-lab/MedSAM)
- [Segment anything in medical images](https://www.nature.com/articles/s41467-024-44824-z)
- [Whisper](https://github.com/openai/whisper)
- [Introducing Whisper](https://openai.com/index/whisper/)
- [Seamless Communication](https://github.com/facebookresearch/seamless_communication)
- [SeamlessM4T: Massively Multilingual & Multimodal Machine Translation](https://arxiv.org/abs/2308.11596)
- [StarCoder2](https://github.com/bigcode-project/starcoder2)
- [StarCoder 2 and The Stack v2: The Next Generation](https://arxiv.org/abs/2402.19173)
- [OpenAI open-weight models (gpt-oss)](https://help.openai.com/en/articles/11870455)
- [SmolVLM: Redefining small and efficient multimodal models](https://huggingface.co/blog/smolvlm)
- [Gemma 3n model card](https://ai.google.dev/gemma/docs/gemma-3n/model_card)
- [Introducing Gemma 3n: The developer guide](https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide/)
- [Stable Diffusion](https://github.com/CompVis/stable-diffusion)
- [Stable Diffusion public release](https://stability.ai/news-updates/stable-diffusion-public-release)
- [AudioCraft](https://github.com/facebookresearch/audiocraft)
- [MusicGen model card](https://github.com/facebookresearch/audiocraft/blob/main/model_cards/MUSICGEN_MODEL_CARD.md)
- [Simple and Controllable Music Generation](https://arxiv.org/abs/2306.05284)
- [Octo: An Open-Source Generalist Robot Policy](https://octo-models.github.io/paper.pdf)
- [Octo generalist robot policy](https://github.com/octo-models/octo)
- [Octo: An Open-Source Generalist Robot Policy](https://arxiv.org/abs/2405.12213)
- [Foundation-Sec-8B model card](https://huggingface.co/RedHatAI/Foundation-Sec-8B)
- [Llama-3.1-FoundationAI-SecurityLLM-Base-8B Technical Report](https://arxiv.org/abs/2504.21039)
- [Anomalib: A Deep Learning Library for Anomaly Detection](https://arxiv.org/abs/2202.08341)
- [Anomalib](https://github.com/open-edge-platform/anomalib)
- [MVTec AD industrial anomaly-detection dataset](https://www.mvtec.com/research-teaching/datasets/mvtec-ad)
- [Narrowing the Gap: Supervised Fine-Tuning of Open-Source LLMs as a Viable Alternative to Proprietary Models for Pedagogical Tools](https://arxiv.org/abs/2507.05305)
- [Qwen3](https://github.com/QwenLM/Qwen3)
- [Cosmos3-Nano model card](https://build.nvidia.com/nvidia/cosmos3-nano/modelcard)
- [Cosmos 3 documentation](https://docs.nvidia.com/cosmos/latest/cosmos3/index.html)
- [NVIDIA Cosmos repository](https://github.com/NVIDIA/cosmos)
- [GPT-2 made release itself part of the debate](https://openai.com/index/gpt-2-1-5b-release/)
- [A community trained GPT-Neo](https://github.com/EleutherAI/gpt-neo)
- [BLOOM paired scale with visible governance](https://huggingface.co/bigscience/bloom)
- [LLaMA accelerated the open-weight ecosystem](https://arxiv.org/abs/2302.13971)
- [Mistral 7B normalized a compact Apache release](https://github.com/mistralai/mistral-src)
- [OLMo made “fully open” a concrete claim](https://allenai.org/olmo2)
- [DeepSeek-R1 released reasoning weights](https://github.com/deepseek-ai/DeepSeek-R1)
- [Hugging Face Hub](https://huggingface.co/docs/hub/index)
- [MLX](https://github.com/ml-explore/mlx)
- [Ollama](https://github.com/ollama/ollama)
- [vLLM](https://github.com/vllm-project/vllm)
- [SGLang](https://github.com/sgl-project/sglang)
- [Transformers](https://github.com/huggingface/transformers)
- [PEFT](https://github.com/huggingface/peft)
- [LM Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness)
- [Open WebUI](https://github.com/open-webui/open-webui)

## Continue exploring

- [Qwen3.8 — weights, license and what is actually open](https://brightaifuture.com/open-models/qwen)
- [DeepSeek V4.1 Flash — weights, license and what is actually open](https://brightaifuture.com/open-models/deepseek)
- [Mistral Small 4 — weights, license and what is actually open](https://brightaifuture.com/open-models/mistral)
- [Gemma 4 — weights, license and what is actually open](https://brightaifuture.com/open-models/gemma)
- [gpt-oss — weights, license and what is actually open](https://brightaifuture.com/open-models/gpt-oss)
- [Llama 4 — weights, license and what is actually open](https://brightaifuture.com/open-models/llama)
- [Olmo 3 — weights, license and what is actually open](https://brightaifuture.com/open-models/olmo)
- [Nemotron 3 Nano Omni — weights, license and what is actually open](https://brightaifuture.com/open-models/nemotron)
- [SmolLM3 — weights, license and what is actually open](https://brightaifuture.com/open-models/smollm)
- [Molmo 2 — weights, license and what is actually open](https://brightaifuture.com/open-models/molmo)
- [Kimi K2.5 — weights, license and what is actually open](https://brightaifuture.com/open-models/kimi)
- [GLM-5.3 Flash — weights, license and what is actually open](https://brightaifuture.com/open-models/glm)
- [Evo 2 — weights, license and what is actually open](https://brightaifuture.com/open-models/evo)
- [Isaac GR00T — weights, license and what is actually open](https://brightaifuture.com/open-models/groot)
- [Voxtral Realtime — weights, license and what is actually open](https://brightaifuture.com/open-models/voxtral)
- [Forecasting the atmosphere ten days ahead](https://brightaifuture.com/discoveries/graphcast)
- [An AI forecast enters daily operations](https://brightaifuture.com/discoveries/aifs)
- [Weather models that can be carried into local workflows](https://brightaifuture.com/discoveries/nvidia-earth-2)
- [Teaching a robot from demonstrations](https://brightaifuture.com/discoveries/openvla-robot-manipulation)
- [A shared starting point for humanoid skills](https://brightaifuture.com/discoveries/groot-humanoid-skills)
- [Predicting a protein’s shape from its sequence](https://brightaifuture.com/discoveries/esmfold-protein-structures)
- [Assembling biological model workflows](https://brightaifuture.com/discoveries/bionemo-biological-workflows)
- [Proposing materials for a desired property](https://brightaifuture.com/discoveries/mattergen-material-candidates)
- [Adapting one Earth model to many maps](https://brightaifuture.com/discoveries/prithvi-earth-observation)
- [SAM 1: selecting an object with a point or a box](https://brightaifuture.com/discoveries/segment-anything-annotation)
- [Giving medical-image annotators an editable first boundary](https://brightaifuture.com/discoveries/medsam-medical-segmentation)
- [Transcribing speech on infrastructure you control](https://brightaifuture.com/discoveries/whisper-local-transcription)
- [Moving between speech and text across languages](https://brightaifuture.com/discoveries/seamlessm4t-speech-translation)
- [Adapting a code model to a private codebase](https://brightaifuture.com/discoveries/starcoder2-local-code)
- [Running a reasoning model on your own machine](https://brightaifuture.com/discoveries/gpt-oss)
- [Asking questions about an image on a small device](https://brightaifuture.com/discoveries/smolvlm-on-device-vision)
- [Combining image, audio, video, and text at the edge](https://brightaifuture.com/discoveries/gemma-3n-edge-assistance)
- [Adapting an image model to a visual practice](https://brightaifuture.com/discoveries/stable-diffusion-adapted-creation)
- [Sketching music from words and melody](https://brightaifuture.com/discoveries/musicgen-conditioned-music)
- [Adapting one robot policy across nine platforms](https://brightaifuture.com/discoveries/octo-robot-policy-adaptation)
- [A security model for a SOC's own evidence](https://brightaifuture.com/discoveries/foundation-sec-8b)
- [A local visual-inspection workbench](https://brightaifuture.com/discoveries/anomalib-industrial-inspection)
- [Explaining a novice programmer's compiler error](https://brightaifuture.com/discoveries/open-model-pedagogy-compiler-errors)
- [Simulating a scene before a robot enters it](https://brightaifuture.com/discoveries/cosmos-3-physical-ai-prototyping)
- [Open intelligence immersive experience](https://brightaifuture.com/open-intelligence)
- [Open thread](https://brightaifuture.com/threads/open)