built throughORANGEBOX·see what it ships·$1 →
An open dark leather folio with a bio-cyan bookmark — open-weight models are the field's library.

AtomEons / Learn / open-weights

The open-weights model index

What you can actually download, fine-tune, and ship — license-honest, hype-free

Open-weights is not open-source. It is a narrower, more honest term: the trained parameters are downloadable, but the training data, training code, alignment recipe, and sometimes the right to use the outputs for commercial purposes — those vary release by release. This page is a reference index of the notable open-weights families as of mid-2026 best-effort. We list parameter counts, native context windows, license families, what each model is meant for, and where the license has teeth. We do not rank by leaderboard. Leaderboards rotate. Licenses persist. A model that scores two points higher on MMLU is irrelevant if its license bars your use case, and a 7B model under Apache 2.0 will out-ship a 70B model under a custom community license for most teams who actually have to deploy something. So the structure here is: index first (params, context, license, base, recommended use), then a plain-language section on what the licenses actually let you do, then a short timeline of how we got here. Where a fact is moving fast — pricing on hosted endpoints, exact benchmark numbers, whether a specific Llama 4 variant has shipped — we mark it as a best-effort snapshot and point you to the provider's official model card or license file. Treat those as the source of truth. Treat this page as the map, not the territory. A note on what is deliberately excluded: closed-weights commercial APIs (GPT-4-class, Claude, Gemini Pro) are not on this list even when they are excellent, because you cannot download them. Models released only via gated waitlists with no actual weights download are also excluded. The bar for inclusion is: weights publicly downloadable from Hugging Face or an equivalent registry, under a license that permits at least research use, with a model card that names the parameter count.

The index — open-weights families worth knowing

FamilyLlama 3.1 (Meta)
Sizes (params)8B · 70B · 405B
Native context128k
License familyLlama 3.1 Community License (custom · permissive with caveats)
Recommended useGeneral-purpose assistant, fine-tune base, English-heavy
FamilyLlama 3.2 (Meta)
Sizes (params)1B · 3B (text) · 11B · 90B (vision)
Native context128k
License familyLlama 3.2 Community License (adds restrictions on EU vision use)
Recommended useOn-device (1B/3B), multimodal (11B/90B)
FamilyMistral Large 2 (Mistral AI)
Sizes (params)123B
Native context128k
License familyMistral Research License (research only · commercial via Mistral)
Recommended useResearch baseline, evaluation, distillation source
FamilyMixtral 8x7B / 8x22B (Mistral AI)
Sizes (params)47B / 141B active-routed MoE
Native context32k / 64k
License familyApache 2.0
Recommended useCommercial deployment, mixture-of-experts research
FamilyCodestral (Mistral AI)
Sizes (params)22B
Native context32k
License familyMistral Non-Production License (research/eval only — commercial requires license)
Recommended useCode-completion research; do not ship commercially without a Mistral commercial license
FamilyDeepSeek-V3 (DeepSeek)
Sizes (params)671B total · ~37B active (MoE)
Native context128k
License familyDeepSeek License (permissive, commercial allowed with use restrictions)
Recommended useStrong general + code; cost-efficient inference
FamilyDeepSeek-R1 (DeepSeek)
Sizes (params)671B total · ~37B active (MoE)
Native context128k
License familyMIT License (model weights)
Recommended useReasoning research; the first widely-distributed open reasoning model
FamilyQwen 2.5 (Alibaba)
Sizes (params)0.5B · 1.5B · 3B · 7B · 14B · 32B · 72B
Native context128k (some variants 32k)
License familyApache 2.0 (most sizes) · Qwen License (72B)
Recommended useMultilingual (esp. Chinese-English), code variants strong
FamilyGemma 2 (Google)
Sizes (params)2B · 9B · 27B
Native context8k
License familyGemma Terms of Use (custom permissive, prohibits certain use cases)
Recommended useEfficient inference, research distillation
FamilyPhi-3 / Phi-3.5 (Microsoft)
Sizes (params)3.8B (mini) · 7B (small) · 14B (medium) · MoE 41B
Native context4k–128k by variant
License familyMIT License
Recommended useSmall-model strength, on-device, fine-tune base
FamilyPhi-4 (Microsoft)
Sizes (params)14B
Native context16k
License familyMIT License
Recommended useReasoning-tuned dense model, late-2024 release
FamilyGranite 3 (IBM)
Sizes (params)2B · 8B (dense) · 1B · 3B MoE
Native context4k–128k
License familyApache 2.0
Recommended useEnterprise-friendly licensing, code variants, time-series variant
FamilyYi 1.5 (01.AI)
Sizes (params)6B · 9B · 34B
Native context4k–200k by variant
License familyApache 2.0 (weights) · Yi License (some commercial terms)
Recommended useBilingual (Chinese-English), long-context variants
FamilyFalcon 2 / Falcon 3 (TII)
Sizes (params)11B · 1B–10B (Falcon 3)
Native context8k–32k
License familyTII Falcon License 2.0 (permissive, includes acceptable-use policy)
Recommended useUAE-hosted base model, multilingual
FamilyStable LM 2 (Stability AI)
Sizes (params)1.6B · 12B
Native context4k
License familyStability AI Community License (membership tiering for commercial use)
Recommended useResearch, small-model experimentation
FamilyOLMo / OLMo 2 (AI2)
Sizes (params)1B · 7B · 13B
Native context4k
License familyApache 2.0 (weights, code, data, recipe all open)
Recommended useReproducible research — the most truly open of the batch
FamilyPythia (EleutherAI)
Sizes (params)70M · 160M · 410M · 1B · 1.4B · 2.8B · 6.9B · 12B
Native context2k
License familyApache 2.0
Recommended useInterpretability research, training-dynamics studies
FamilyBLOOM (BigScience)
Sizes (params)560M · 1.1B · 1.7B · 3B · 7.1B · 176B
Native context2k
License familyBigScience RAIL License v1.0 (Responsible AI license with use-case restrictions)
Recommended useMultilingual baseline (46 natural + 13 programming languages), historical reference

What "open weights" actually means

Open weights means you can download the trained model parameters. That is the minimum. Beyond that, every claim about "openness" has fine print, and the fine print is where you get burned. There are roughly four tiers of openness in practice. First, fully open: weights, training code, training data, and the training recipe are all published under a permissive license. OLMo from AI2 is the canonical example — they publish the data mix (Dolma), the training code (OLMo-core), the checkpoints, and the evaluation suite. Pythia from EleutherAI similarly publishes intermediate checkpoints for interpretability work. Second, open weights with open code: the model file is downloadable, the inference code is on GitHub, but the training data and recipe are not disclosed. Mixtral 8x7B, Qwen 2.5, most Granite releases sit here. Third, open weights under a custom community license: Meta's Llama 3.x family, Google's Gemma 2, Stability's models. The weights download, but the license adds use restrictions — sometimes geographic (Llama 3.2's vision models are restricted in the EU for the model provider, not the deployer), sometimes user-count gated (Llama's >700M MAU clause requires a separate commercial license), sometimes acceptable-use clauses that prohibit specific verticals. Fourth, research-only weights: Mistral Large 2 and Codestral fall here. The weights are downloadable from Hugging Face but the license says non-commercial / research-only, and commercial use requires a paid license from Mistral. The practical lesson: read the LICENSE file. Not the marketing page. Not the README. The LICENSE file in the model repo. If it says "Llama 3.1 Community License," pull up Meta's PDF and search it for your use case. If it says "Mistral Research License," you cannot ship a paid product on it. If it says "Apache 2.0" or "MIT," you have the broadest latitude, but you still owe attribution and you should still read it.

License gotchas worth flagging

Three landmines that catch teams repeatedly. (1) The Llama Community License has a clause that triggers when a deploying entity has more than 700 million monthly active users — at that point you owe Meta a separate license. Below that ceiling, commercial use is permitted with attribution. Most teams are fine. Some are not, and the threshold is per-organization, not per-product. (2) Distillation: many community licenses restrict using the model's outputs to train a competing foundation model. This means generating a synthetic dataset with Llama 3 and using it to train a from-scratch competitor is contractually prohibited under Meta's license, even though the outputs themselves are not copyrighted. (3) Acceptable use policies: Llama, Gemma, and Falcon all attach AUPs that prohibit certain applications (weapons of mass destruction, large-scale surveillance, content sexualizing minors, etc.). These are enforceable contract terms, not just guidelines. Apache 2.0 and MIT models have no such restrictions in the license itself, but you are still bound by applicable law.

Frontier open-weights families — the four to actually pay attention to

Llama (Meta)

Llama 3.1 · Llama 3.2 · Llama 3.3 70B

The most-deployed open-weights family. Strong English performance, broad fine-tune ecosystem, 128k context on the 3.1/3.3 generation, multimodal vision variants in 3.2. License is custom (Llama Community License), not Apache, and the 700M MAU clause matters at scale. Llama 4 has not, as of June 2026 best-effort, shipped to public weights — check Meta's official announcements. If you are choosing a default open-weights base for an English-heavy product and you are below the MAU threshold, Llama 3.x is the boring correct answer.

DeepSeek (DeepSeek-AI)

DeepSeek-V3 · DeepSeek-R1 · DeepSeek-Coder-V2

Mixture-of-experts architecture with 671B total parameters and roughly 37B activated per token, which gives strong quality at inference cost closer to a dense ~70B. R1 was the first widely-distributed open-weights reasoning model and was released under MIT license — an unusually permissive choice. V3 ships under DeepSeek's own license with commercial use permitted under documented restrictions. Strong on code and math. Hosted inference is unusually cheap; self-hosting the full MoE requires serious GPU memory.

Qwen (Alibaba)

Qwen 2.5 · Qwen 2.5-Coder · Qwen 2.5-VL

Most of the Qwen 2.5 lineup ships Apache 2.0 (excluding the 72B variant which uses the Qwen License with commercial-use terms). Strong multilingual performance, especially Chinese-English, and the Coder variants are competitive with closed models on HumanEval and similar code benchmarks. Range from 0.5B (on-device) through 72B. Good default for a multilingual product or a small-model experiment.

Mistral (Mistral AI)

Mixtral 8x7B / 8x22B · Mistral Large 2 · Codestral · Mistral 7B

Bifurcated portfolio: the older Mixtral models and Mistral 7B are Apache 2.0 and freely usable commercially. The newer Mistral Large 2 and Codestral are research-license-only — downloadable from Hugging Face for evaluation, but commercial deployment requires a paid Mistral license. Mistral was the first major lab to popularize MoE in open weights. Strong European data coverage.

Small-model picks — under 14B parameters

ModelLlama 3.2 1B / 3B
Params1B · 3B
LicenseLlama 3.2 Community
Notable propertyTuned for on-device; 128k context retained
ModelPhi-3.5-mini
Params3.8B
LicenseMIT
Notable propertyPunches well above weight on reasoning benchmarks for size
ModelPhi-4
Params14B
LicenseMIT
Notable propertyMicrosoft's late-2024 dense reasoning model
ModelGemma 2 2B / 9B
Params2B · 9B
LicenseGemma Terms
Notable propertyTight inference footprint, 8k context
ModelQwen 2.5 0.5B / 1.5B / 3B / 7B
Params0.5B–7B
LicenseApache 2.0
Notable propertyMultilingual, strong code variants
ModelGranite 3 2B / 8B
Params2B · 8B
LicenseApache 2.0
Notable propertyEnterprise-clean license, IBM-backed
ModelOLMo 2 7B / 13B
Params7B · 13B
LicenseApache 2.0 (fully open)
Notable propertyReproducible research-grade release
ModelMistral 7B (v0.3)
Params7.3B
LicenseApache 2.0
Notable propertyBattle-tested fine-tune base, broad ecosystem

How we got here — a short timeline

  1. 2022-07

    BLOOM (BigScience)

    176B-parameter multilingual model from the BigScience workshop coordinated by Hugging Face. First model of this scale released under a Responsible AI License with explicit use-case restrictions. Set a baseline that openness could include conditions.

  2. 2023-02

    LLaMA 1 (Meta · research-only leak)

    Meta releases LLaMA 1 under a research-only license. Weights leak via 4chan within a week. The leak forces the industry to confront the gap between intended access and actual access, and arguably accelerates the next year of open releases.

  3. 2023-07

    Llama 2

    Meta releases Llama 2 with a custom Community License permitting commercial use below 700M MAU. This is the moment open-weights becomes commercially viable at scale — the entire fine-tune ecosystem (Vicuna, WizardLM, Alpaca derivatives, etc.) coalesces around it.

  4. 2023-09

    Mistral 7B

    Mistral AI ships a 7B dense model under Apache 2.0. Outperforms Llama 2 13B on many benchmarks and ignites the small-model serious-quality thread that still runs.

  5. 2023-12

    Mixtral 8x7B

    Mistral releases the first widely-adopted open MoE. Apache 2.0. Demonstrates that mixture-of-experts can be shipped openly, not just kept behind closed APIs.

  6. 2024-04

    Llama 3 · Phi-3 · OLMo

    Meta ships Llama 3 8B/70B with 8k context. Microsoft ships Phi-3 demonstrating small-model strength via curated data. AI2 ships OLMo with full data and training-code disclosure. The 'open weights' field stratifies into permissive-but-closed-data, permissive-with-AUP, and fully-open-everything tiers.

  7. 2024-07

    Llama 3.1 · 405B

    Meta releases Llama 3.1 8B/70B/405B with 128k context. The 405B variant is the largest openly-released dense model to date and shifts the conversation about what 'frontier-class open weights' looks like.

  8. 2024-12

    DeepSeek-V3 · R1 (early 2025)

    DeepSeek releases V3 in December 2024, followed by R1 in January 2025 under MIT license. R1 is the first widely-distributed open-weights reasoning model and the release shifts both the cost-curve and the political conversation around open-weights releases.

  9. 2025-2026

    Continued releases — best-effort tracking

    Qwen, Mistral, Meta, DeepSeek, IBM, Microsoft, AI2, and 01.AI continue iterating. Llama 4 timing and Llama 3.3 70B (released late 2024) are the moving pieces — check the provider model cards for current state. This page reflects what is verifiably released as of mid-2026 best-effort.

Choosing a base — a short decision tree

  • You need maximum permissive license (Apache 2.0 or MIT) and English-heavy use: Mistral 7B, Mixtral 8x7B, Phi-3.5, Granite 3, OLMo 2, Qwen 2.5 (most sizes).
  • You need maximum permissive license and multilingual coverage (incl. Chinese): Qwen 2.5 (Apache 2.0 sizes) or Yi 1.5.
  • You need 128k context and you are below 700M MAU: Llama 3.1 8B / 70B / 405B or Llama 3.3 70B.
  • You need on-device inference under 4B parameters: Llama 3.2 1B/3B, Phi-3.5-mini, Gemma 2 2B, Qwen 2.5 0.5B / 1.5B / 3B.
  • You need open reasoning model with chain-of-thought: DeepSeek-R1 (MIT-licensed weights) — distill to a smaller dense model for production.
  • You need code-specific: Qwen 2.5-Coder, DeepSeek-Coder-V2, or Codestral (research-only — buy a Mistral license for prod).
  • You need full reproducibility (data + code + weights): OLMo 2 or Pythia. Nothing else in the list publishes the data mix at OLMo's level.
  • You are a >700M MAU company and want a frontier open base: Apache-2.0 path (Mixtral 8x22B, Qwen 2.5 72B under Qwen License with commercial terms, or DeepSeek-V3 under its own commercial-permissive license) or negotiate directly with Meta.

Numbers we deliberately did not invent

We did not publish specific MMLU, GPQA, or HumanEval scores in this index. Those numbers move with each fine-tune, each eval harness version, and each prompt-formatting choice — comparing 'official' numbers across providers using different harnesses is misleading. For current benchmark performance, consult: the model card on Hugging Face, the provider's release post, and an independent harness run (lm-evaluation-harness from EleutherAI is the most-cited). The official Hugging Face Open LLM Leaderboard was retired in mid-2024; community-maintained successors exist but rotate. Treat any single leaderboard as one data point, not the verdict.

Honest limits of this page

This page is a snapshot. Open-weights releases happen monthly. By the time you read this, at least one of these families will have shipped a new variant, one license may have been updated (Meta has updated the Llama Community License between 2 and 3 and 3.2), and at least one provider may have switched between research-only and commercial-permissive (or the reverse). The structural claims — what 'open weights' means versus 'open source,' what license tiers exist, why distillation clauses matter — those are durable. The specific row contents are best-effort as of mid-2026 and you should verify each license against the LICENSE file in the model's actual repo before you build on it. If a fact on this page contradicts a provider's official model card, the official model card wins. If a license clause on this page contradicts the LICENSE file in the repo, the LICENSE file in the repo wins. If we missed a model family that should be on this list, that is on us — the bar is publicly-downloadable weights with a real license, and we tried to cover the families that show up most in production deployments and academic citation as of the snapshot date.

Sources

  1. [01]

    Meta Llama 3.1 release post documents the 8B / 70B / 405B sizes and 128k context window.

    ai.meta.com/blog/meta-llama-3-1/

  2. [02]

    The Llama 3.1 Community License contains the 700 million monthly active users threshold for separate commercial licensing.

    llama.meta.com/llama3_1/license/

  3. [03]

    Meta Llama 3.2 release post documents the 1B / 3B text and 11B / 90B vision variants and the EU restriction for multimodal use.

    ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/

  4. [04]

    Mistral AI announces Mistral Large 2 at 123B parameters with 128k context under the Mistral Research License.

    mistral.ai/news/mistral-large-2407/

  5. [05]

    Mistral AI announces Mixtral 8x7B under Apache 2.0 license as the first widely-adopted open MoE.

    mistral.ai/news/mixtral-of-experts/

  6. [06]

    Mistral AI ships Codestral 22B under the Mistral Non-Production License restricting commercial deployment.

    mistral.ai/news/codestral/

  7. [07]

    DeepSeek-V3 technical report documents the 671B-parameter MoE architecture with roughly 37B activated parameters per token.

    arxiv.org/abs/2412.19437

  8. [08]

    DeepSeek-R1 paper documents the open reasoning model release under MIT license.

    arxiv.org/abs/2501.12948

  9. [09]

    Alibaba's Qwen 2.5 release post documents the 0.5B through 72B size range and the Apache 2.0 license for most variants.

    qwenlm.github.io/blog/qwen2.5/

  10. [10]

    Gemma Terms of Use document the custom permissive license with prohibited-use clauses for Google's Gemma models.

    ai.google.dev/gemma/terms

  11. [11]

    Microsoft Phi-3 technical report documents the 3.8B / 7B / 14B Phi-3 family release under MIT license.

    arxiv.org/abs/2404.14219

  12. [12]

    Microsoft Phi-4 technical report documents the 14B dense reasoning-tuned model.

    arxiv.org/abs/2412.08905

  13. [13]

    IBM Granite 3 model family is released under Apache 2.0 with enterprise-targeted licensing terms.

    ibm.com/granite

  14. [14]

    01.AI's Yi paper documents the 6B / 9B / 34B Yi family and bilingual training corpus.

    arxiv.org/abs/2403.04652

  15. [15]

    TII's Falcon 2 release documents the 11B model under the TII Falcon License 2.0.

    falconllm.tii.ae/falcon-2.html

  16. [16]

    Stability AI's Stable LM 2 release documents the 1.6B and 12B variants under the Stability AI Community License.

    stability.ai/news/introducing-stable-lm-2

  17. [17]

    AI2 OLMo releases publish weights, training data (Dolma), training code, and recipe under Apache 2.0 for full reproducibility.

    allenai.org/olmo

  18. [18]

    EleutherAI Pythia paper documents the 70M through 12B suite released with intermediate checkpoints for interpretability research under Apache 2.0.

    arxiv.org/abs/2304.01373

  19. [19]

    BLOOM paper documents the 176B multilingual model released under the BigScience RAIL License v1.0.

    arxiv.org/abs/2211.05100

  20. [20]

    Hugging Face documents the retirement of the original Open LLM Leaderboard.

    huggingface.co/blog/open-llm-leaderboard-archive

  21. [21]

    EleutherAI's lm-evaluation-harness is the most widely-cited independent benchmark harness for open-weights LLM evaluation.

    github.com/EleutherAI/lm-evaluation-harness

  22. [22]

    The Open Source Definition (OSI) describes the conditions for software to be considered open-source, which most current open-weights model licenses do not meet.

    opensource.org/osd

LAB · ATOMEONS · MARCO ISLAND FLÆONS RESEARCH · 12 PAPERS · CC-BY 4.0ORANGEBOX v1.0.0-beta · TURBO-OPTIMIZE CLAUDE · SHIPPED 2026-05-30B00KMAKR v3.2.0 · AI PUBLISHING COCKPIT · MAC + WINDOWSFREE LAUNCH WEEK · ENDS JUNE 6 · §4A NO-SAAS LOCKFOUNDER'S VIEW · NEXT BROADCAST IN ...CITE THE WORK · FORWARD THE LINK · NO ALGORITHMLAB · ATOMEONS · MARCO ISLAND FLÆONS RESEARCH · 12 PAPERS · CC-BY 4.0ORANGEBOX v1.0.0-beta · TURBO-OPTIMIZE CLAUDE · SHIPPED 2026-05-30B00KMAKR v3.2.0 · AI PUBLISHING COCKPIT · MAC + WINDOWSFREE LAUNCH WEEK · ENDS JUNE 6 · §4A NO-SAAS LOCKFOUNDER'S VIEW · NEXT BROADCAST IN ...CITE THE WORK · FORWARD THE LINK · NO ALGORITHM