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.
01 / OPEN IS MORE THAN A DOWNLOAD
Nine doors. Which ones actually open?
Weights let you run a model. Code, data and permissions determine how far you can take it. These are different kinds of access, not a score.
Olmo 3 7B and 32B · Apache-2.0 for models and code; data components retain their stated licenses. These are source checks, not an OSI certification.
01WeightsavailableInspect +
Downloadable learned parameters for local inspection, adaptation, or inference.
Available = the named artifact is published. Partial = some pieces or conditions. Restricted = material terms limit access or use. Not established = this review did not verify it; that is not evidence of absence.
A vocabulary for seeing clearly · 12 definitionsWeights
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 ↗
02 / THE OPEN MODEL ATLAS
Many starting points. No single winner.
15 curated family records. Sources checked 2026-09-19. Positions are editorial, not rankings, market share or a performance map.
15 families to explore. All 15 checked family records, without a ranking.
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.
The selected checkpoint is stated on every profile. A family can contain permissive and restrictive releases. These source checks are a dated edition, not a claim that every emerging release has been found.
03 / WHAT CAN I RUN?
Some of this can live on your machine.
Start with what you have. Discover a plausible scale, then look at what a runtime actually needs.
Begin with the machine in front of you
What are you using?
Choose a device and the memory available to it. Bright will compare that planning budget with idealized weight storage—not certify that a checkpoint will run.
Mac9.6 GBweight-only budget
40% left for everything beyond weights
Apple silicon uses one physical memory pool shared by the CPU and GPU. The operating system and runtime still need part of it.
In Apple silicon, the CPU and GPU access the same physical memory pool. That differs from a discrete GPU with separate VRAM. Apple MLX explanation ↗
The reveal
Some documented weights enter the realm of plausible.
6 documented checkpoints fall under a 9.6 GB weight-only budget at 4-bit. A supported quantization, runtime and acceptable speed still need to be verified.
This is a planning lens, not a compatibility test. Decimal GB = total parameters × bits ÷ 8. The calculation includes inactive experts but omits packaging overhead, auxiliary modules, activations, runtime memory and context cache. It does not assert that a quantized artifact or supported backend exists.
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.
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.
02
Keep it working
A place without a reliable connection
What changes?
After downloading the necessary artifacts, supported local inference can work offline. Tools that fetch information or call services still need a network.
03
Make it specific
A language, a discipline, a task
What changes?
Fine-tunes and adapters let builders change behavior. They need suitable data, rights, evaluation and a license that permits the intended use.
04
Keep a choice
An institution choosing its infrastructure
What changes?
Self-hosting can give companies, researchers and countries operational control and portability. Hardware, expertise and supply chains remain dependencies.
05
Look inside
A researcher asking why it failed
What changes?
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.
24 documented examples of released models and the work around them. Research artifacts, operational systems and possible uses are labeled separately. Historical examples are not presented as new releases.
01 / 4 documented examplesEarthRead a changing planet.Atmospheres · weather · observationEnter this world ＋
Climate · DemonstratedForecasting 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.
Climate · DeployedAn 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.
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.
Earth · EmergingWeather 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.
Earth · DemonstratedAdapting 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.
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.
02 / 4 documented examplesLifeLook inside living systems.Proteins · biology · medical images · materialsEnter this world ＋
Biology · DemonstratedPredicting 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.
Biology · EmergingAssembling biological model workflows＋
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.
Health · ExperimentalGiving 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.
Materials · DemonstratedProposing 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.
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 · ExperimentalTeaching 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.
Robotics · ExperimentalA 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.
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.
Robotics · DemonstratedAdapting 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 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.
Manufacturing · ExperimentalA 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.
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.
Robotics · ExperimentalSimulating 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.
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 · DemonstratedTranscribing 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.
Access · EmergingMoving between speech and text across languages＋
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.
Culture · ExperimentalSketching 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.
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.
05 / 4 documented examplesCreationSee an image as something editable.Images · vision · multimodal toolsEnter this world ＋
Infrastructure · DemonstratedSAM 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.
Access · EmergingAsking 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.
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.
Access · EmergingCombining 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.
Culture · DemonstratedAdapting 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.
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.
06 / 4 documented examplesWorkBring a model into a discipline.Code · reasoning · security · educationEnter this world ＋
Infrastructure · EmergingAdapting 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.
Open intelligence · EmergingRunning 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.
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.
Infrastructure · EmergingA 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 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.
Education · DemonstratedExplaining 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.
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.
“Open” and “hosted” are not opposites: open-weight models can be hosted by someone else. Compare the model’s access terms and the deployment arrangement separately.
Decision questions, not measured rankings. Derived from the model cards and runtime documentation linked in this guide.
Situation
Run it yourself
Use a hosted service
Ask before deciding
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.
08 / STATE OF OPEN
A record worth returning to.
Dated developments from Bright’s evidence ledger. This edition is curated; nothing here is an automated live feed. The atlas above names newer checked releases where the historical ledger has not yet caught up.
NASA and IBM released a Lunar Foundation Model trained primarily on Lunar Reconnaissance Orbiter data for research tasks such as crater and volcanic-feature mapping and possible polar-ice analysis.
Emerging · Repository licenses must be checked separately for code, weights, and data before calling the whole package open source.
Hugging Face and Voice Arena added Hindi and Indian English evaluation to an open speech-recognition leaderboard with held-out/private splits and demographic and geographic test design.
Emerging · A benchmark expansion is not proof that any product is equitable.
TESSERA provides 10-meter, pixel-wise annual embeddings learned from Copernicus Sentinel-1 and Sentinel-2. ESA says the model first launched in 2025; its repository and a peer-reviewed CVPR 2026 paper now document the work.
Demonstrated · An embedding discards information and is not a universal map.
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.
Experimental · 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.
Access changed. Then what people could build changed.
Selected historical milestones. This is a history of different release choices, not a claim that every release met the same definition of open.
GPT-2 made release itself part of the debate
What changed?
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.
Hugging Face Hub
What they make possible
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.
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.
