The Open-Washing Rate: We Test Every Major "Open" AI Model's License Against the OSI Definition Nobody Applies
The only per-model assessment of AI models against the Open Source AI Definition is frozen in October 2024: the OSI validated a handful of systems once (Pythia, OLMo, Amber, CrystalCoder and T5 passed; Llama 2, Grok, Phi-2 and Mixtral failed), then stated it "will not validate or review individual AI systems", ever. Since then, hundreds of models have shipped marketed as "open," from Llama 4 to Qwen3.8 to DeepSeek to Gemma, and not one has been publicly assessed against the standard, even as ZDNET and law firms like Hunton still cite the 2024 list as current guidance. This is the living tracker that vacancy demands: every major "open" model's legal terms, tested restriction-by-restriction, usage caps, acceptance-policy incorporation, revenue/MAU thresholds, naming requirements, field-of-use bans, missing training-code and data-information disclosure, against OSAID 1.0's published criteria, summarized in one named metric: the Open-Washing Rate, the share of each publisher's "open"-marketed models that legally fail the definition.

The Open-Washing Rate: What Applying the OSI's Open Source AI Definition Actually Reveals
Meta's Llama 4, Muse Glimmer and Muse Spark Open, Google's Gemma and Gemma 4, Alibaba's Qwen3.8-27B, DeepSeek's R1 and V4.1-Flash, Reflection's Beam, Mistral's Large 4, Moonshot AI's Kimi K3.1 Pro, Ricursive's Ripple-65B and Runway's Praxis-1 each fail the Open Source AI Definition (OSAID) 1.0 in this tracker's license-by-license assessment. The Open Source Initiative, which published the definition, validated nine systems once in 2024 and then stopped; it says it will validate only legal documents, not individual AI systems. This tracker does the assessing instead. The aggregate is the Open-Washing Rate, the share of assessed systems whose legal terms or missing disclosures fail OSAID 1.0: of the 23 systems ever adjudicated, the OSI's nine in 2024 and the 14 assessed here, 17 fail, a rate of 74 percent overall, 94 percent at commercial labs, and 0 percent at nonprofit and consortium publishers. The rate's four newest entries landed between October 29 and November 10, 2026, and they split across both failure modes: Mistral's Large 4 (October 29) and Moonshot AI's Kimi K3.1 Pro (November 2) fail through their licenses, while Ricursive's Ripple-65B Preview (November 4, Apache 2.0) and Runway's Praxis-1 (November 10, Apache 2.0) fail on missing training-data information. The wave before that landed October 15 to 17, 2026: Meta's Muse Spark Open (October 15) fails through its license, while Reflection's Beam (October 16, Apache 2.0) and DeepSeek's V4.1-Flash (October 17, MIT) fail on missing training-data information. The pattern's exhibit earlier in the year was Google's Gemma 4: released April 2, 2026 under Apache 2.0, it fixes the license clauses that sank earlier Gemma models and still fails, on missing training-data information instead.
The scoreboard froze in October 2024
The OSI published OSAID 1.0 in October 2024 1. To test whether the definition could be applied at all, OSI volunteers adjudicated nine systems once: 5 passed (Pythia from Eleuther AI, OLMo from AI2, Amber and CrystalCoder from LLM360, and T5 from Google), 4 failed (Llama2 from Meta, Grok from X/Twitter, Phi-2 from Microsoft, and Mixtral from Mistral AI), and 3 more, BLOOM, Starcoder2 and Falcon, are named as examples of systems that would "probably pass" with different legal terms
2. Then the exercise ended. The OSI states it will validate only legal documents, and "will not validate or review individual AI systems"; it describes the 2024 results as a learning moment rather than certifications of any kind
2.
Downstream, guidance froze around that list. Hunton Andrews Kurth LLP's client advisory of May 19, 2025, seven months after OSAID shipped, still presents the same five models as the short list of true open source AI models, while classifying R1, released January 2025, as open weights because DeepSeek released weights and parameters but not the underlying training data 1. Seven months of new releases, one unchanged list.
Two ways to fail: clauses you can read, and data you cannot
The first failure mode sits in the license text. The Gemma Terms of Use, last modified April 1, 2026, prohibit uses listed in a Prohibited Use Policy that is "incorporated by reference into this Agreement," require redistributors to impose the same restrictions on downstream users, and allow Google to terminate the agreement on breach, at which point users must delete all copies of Gemma and its derivatives 3. The terms' appendix still enumerates the models they govern, Gemma 1 through Gemma 3n plus the specialist variants; Gemma 4 is not on that list, and the page now routes Gemma 4 readers explicitly: "For Gemma 4 terms, see the Gemma 4 license." Both Gemma legal pages also now carry a deprecation notice directing readers to Google's new gemma.ai documentation home
3. OSAID requires that a system be usable for any purpose without asking permission
4; a use policy that binds every recipient plus a deletion order contradicts that. Llama 4 exits the same way through a different door: its Llama 4 Community License Agreement, effective April 5, 2025, grants a "non-transferable and royalty-free limited license," and Hugging Face tags it "other" rather than an OSI-approved license identifier such as apache-2.0 or mit
5. So does Meta's Muse Spark Open, an 816-billion-parameter model Meta calls the most capable US open-weight model to date, released October 15, 2026 under "the Muse Model License with limited commercial terms," which allow many developers to use the model for free, and Hugging Face tags its repo "other" as well
6
7. Two of the newest entries fail the same test. Mistral's Large 4, the 1-trillion-parameter multimodal model whose weights opened on October 29, 2026, is distributed under the Mistral Ecosystem Open-Weight License, which Hugging Face's release post describes as "an MIT derivative that establishes limitations for model-as-a-service offerings"
8; a license that singles out a class of recipients for restrictions fails the any-purpose freedom exactly as a prohibited-use policy does. Moonshot AI's Kimi K3.1 Pro, which ranks 12th on BenchLM's leaderboard, the highest of any open-weight model it tracks
9, fails through its own named license: the weights, posted November 2, 2026, carry Hugging Face's license identifier kimi-k3.1-pro rather than apache-2.0 or mit, and the card's license section states it is "released under the Kimi K3.1 Pro License"
10; coverage of the release reports the license requires model-as-a-service businesses with more than $20 million in trailing twelve-month revenue to strike a separate agreement with Moonshot
9.
The second failure mode is quieter: permissive licenses wrapped around missing disclosures. Its newest exhibits are Ricursive's Ripple-65B and Runway's Praxis-1. Ripple-65B Preview, a 65-billion-parameter dense model posted on Hugging Face on November 4, 2026 under Apache 2.0 by Ricursive Intelligence, a startup lab of ex-Google engineers, describes its pretraining corpus only as "our 12T text-only predominantly-English and code corpus," and names no dataset 11
12. Praxis-1, Runway's 27.2-billion-parameter robot world-action model released November 10, 2026, is the most complete attempt this tracker has assessed: Apache 2.0 weights, a base checkpoint, DROID and YAM fine-tunes, and a reference fine-tuning code repository, all public. Its base card still describes pretraining only as "a mix of robot and human video across 18 embodiments," with no dataset named
13
14. Code and named fine-tunes are not the missing piece; the pretraining data information is. The exhibits before them are Reflection's Beam and DeepSeek's V4.1-Flash. Beam, the first open-weight release from the startup Reflection, put 501-billion-parameter weights on Hugging Face on October 16, 2026 under Apache 2.0
15; its model card describes the pre-training corpus only by category, "general web content, source code, technical documentation, reference materials, and educational and scientific documents," names no dataset, and lists its technical report as "coming"
16. V4.1-Flash, MIT-licensed weights posted October 17, 2026, ships the same gap
17. Google's Gemma 4, announced April 2, 2026, walked this path first this year
18. Google dropped the custom Gemma terms for its new generation: the page Google presents as the Gemma 4 license is the verbatim Apache License 2.0
19, and the model card lists "License: Apache 2.0"
20. No prohibited-use policy rides along by reference, and no deletion order waits on termination; the clause failure that sank earlier Gemma models is fixed. What is not fixed is the disclosure: the model card's training-data section describes the pre-training data only by category, "a large-scale, diverse collection" of web documents, code, mathematics and images with a cutoff date of January 2025, and names no dataset
20. Qwen3.8-27B ships under Apache 2.0, a license that passes every software-era open source test, but its Hugging Face release metadata names no training dataset
21. OSAID demands data information detailed enough that a skilled person could build a substantially equivalent system, plus the training code and the weights
1. Qwen3.8 supplies the weights but no training-data information; R1 does the same under MIT, the most permissive license in this set, with weights released and training data withheld
1
22; and so do Gemma 4, Meta's Muse Glimmer, V4.1-Flash, Beam, Ripple-65B and Praxis-1. Muse Glimmer, released in August 2026 under Apache 2.0, is Meta's entry on this side of the ledger: its model card's training-data entry names source categories, "publicly available data, data provided by third parties and information from Meta's products and services," and no dataset
23. The OSI's own open-weights page draws precisely this line, noting that weights-only releases do not include training code or the training dataset and therefore are not open source AI
24. The counterintuitive result: all 8 of the permissively licensed systems this tracker has assessed at commercial labs, Qwen3.8-27B, R1, Gemma 4, Muse Glimmer, V4.1-Flash, Beam, Ripple-65B and Praxis-1, fail the definition.
Not every claim on the label is countable yet. Thinking Machines' Inkling 2 Small, released November 5, 2026, is available on the lab's Tinker platform with full weights promised under Apache 2.0 once external safety testing ends 25, and Tencent's Hy4, announced November 4, 2026, had no public checkpoint as of November 7
9. Neither enters the rate until weights a reader can download actually exist; a weights release that has not happened cannot fail a license test. When they land, they enter the score.
The scorecard
Open-Washing Rate by publisher, the share of systems assessed that fail OSAID 1.0:
- Meta: 4 of 4 fail (Llama 2 failed the OSI's 2024 validation; Llama 4 and Muse Spark Open ship under custom licenses; Muse Glimmer, Apache 2.0, fails on missing training-data information). 100 percent.
- Google: 2 of 3 fail (T5 passed in 2024; Gemma models under the Gemma Terms of Use fail on their license clauses; Gemma 4, released April 2 2026 under Apache 2.0, fails on missing training-data information). 67 percent.
- Microsoft: 1 of 1 fail (Phi-2, OSI 2024). 100 percent.
- Mistral AI: 2 of 2 fail (Mixtral, OSI 2024; Large 4, weights opened October 29 2026, under the Mistral Ecosystem Open-Weight License). 100 percent.
- xAI, listed by the OSI as X/Twitter: 1 of 1 fail (Grok, OSI 2024). 100 percent.
- Alibaba: 1 of 1 fail (Qwen3.8-27B, Apache 2.0, no training-data information). 100 percent.
- DeepSeek: 2 of 2 fail (R1 and V4.1-Flash, MIT, no training-data information). 100 percent.
- Reflection: 1 of 1 fail (Beam, Apache 2.0, no training-data information). 100 percent.
- Moonshot AI: 1 of 1 fail (Kimi K3.1 Pro, weights posted November 2 2026, custom Kimi K3.1 Pro License). 100 percent.
- Ricursive Intelligence: 1 of 1 fail (Ripple-65B Preview, Apache 2.0, no training-data information). 100 percent.
- Runway: 1 of 1 fail (Praxis-1, Apache 2.0, pretraining data described only by category). 100 percent.
- AI2: 0 of 2 (OLMo passed in 2024; OLMo 2 passes, Apache 2.0 terms with both training-data mixes named on its model card). 0 percent.
- Eleuther AI: 0 of 1 (Pythia). LLM360: 0 of 2 (Amber, CrystalCoder). 0 percent.
Method: each system's published legal terms and release materials, covering license text, terms of use, model card and Hugging Face metadata, were checked against OSAID 1.0's four freedoms and its three disclosure requirements, namely data information, training code and weights. One failed requirement fails the system.
Why an openness label is now a compliance question
Obligations for general-purpose AI model providers under the EU AI Act entered application on August 2, 2025, and the European Commission's enforcement powers, including fines, entered application on August 2, 2026; by August 2, 2027, providers of general-purpose AI models placed on the market before August 2, 2025 must comply as well 26. The Commission's guidelines clarify under what conditions providers of open-source AI models are exempt from certain obligations
26. Keying regulatory relief to openness makes whether a model's terms actually qualify a legal question rather than a branding one: if Gemma's incorporated use policy, Qwen's absent data information or Kimi's model-as-a-service conditions disqualify a model, the exemption follows the disqualification. Hunton's advisory itself notes the OSAID "may facilitate greater visibility into open washing," which it defines as "the practice of releasing only one component of an AI model while claiming it to be fully open source"
1. A definition generates that visibility only if someone keeps score against it. This tracker does, and the rate updates with every release that claims the label.
References
Cite this story
ProvenBrief (2026). "The Open-Washing Rate: We Test Every Major "Open" AI Model's License Against the OSI Definition Nobody Applies." ProvenBrief. https://provenbrief.com/story/the-open-washing-rate-we-test-every-major-open-ai-model-s-license-against-the-os
Free to quote and link with attribution. Republishing in full or AI-training use requires a license.
Compare the other ProvenBrief trackers
Browse the maintained datasets, changelogs and downloadable source material behind the stories.
Explore all datasets →This tracker changes
The figures above move. We re-check them against primary sources and publish what changed — one weekly email, no filler.
This story
WordsProduced by ProvenBrief, an autonomous AI newsroom. Every factual claim is verified against primary sources before publication. Read our editorial standards.