Say you pick an open model to ship inside a product. The word “open” sits right there in the name, so the open weight license feels like a formality you can skip.
But “open” is a marketing word, not a license.
I pulled the 100 most downloaded open models on Hugging Face and checked each one’s license tag and gating status. Only 76 were commercially free to use. Fourteen carried no license tag at all. And the models most of us picture when we hear “open AI,” Meta’s Llama and Google’s Gemma, were gated behind an access request.
Key Takeaways
I measured open weight licenses across the 100 most downloaded models on Hugging Face. Seventy-six were commercially free, while 14 carried no license tag at all. The “open” flagships from Meta and Google were gated behind access approval, and the labs that spread permissive licenses widest were challengers like Qwen and DeepSeek.
- Top 100 by downloads: 76 permissive, 14 undeclared, 6 gated, 4 custom
- Every measured Meta Llama variant and google/gemma-7b required access approval
- Llama’s license demands a separate grant above 700 million monthly active users
- For a commercial product, Apache 2.0 and MIT stay the cleanest picks
Downloads Are Dominated By Free Licenses
Let me start with what I measured and how. From the public Hugging Face API I pulled the top 100 models by download count, then read each model’s license tag and gating flag. I sorted them into four buckets.
Permissive covers Apache 2.0, MIT and similar. Gated means an access request stands between you and the weights. Undeclared means no license tag exists. Custom covers everything with use restrictions. The snapshot is dated July 25, 2026, and the raw output is preserved.
The result was a landslide for permissive licensing. Of the top 100, seventy-six were permissive. By tag frequency, Apache 2.0 appeared 58 times and MIT 19 times. The download charts are carried by embeddings and small utility models, and almost all of them ship under Apache or MIT.
This lines up with my earlier finding that the most downloaded model was an embedding model, not a chatbot. The chart is ruled by small models doing quiet search and classification work rather than the headline LLMs.
| Bucket | Count | Typical license |
|---|---|---|
| permissive (commercial-free) | 76 | Apache 2.0, MIT |
| undeclared | 14 | none |
| gated (access request) | 6 | Llama, pyannote |
| restrictive / custom | 4 | custom terms |
The Models Branded “Open” Are The Closed Ones
The story turns when you isolate the flagships people actually talk about. I sorted text generation models by likes and looked at the top of that list. Meta’s Llama family, in the Llama 2, Llama 3 and Llama 3.1 variants present at measurement time, were all gated behind an access request, and they ship under Meta’s own community license rather than a standard open source one. google/gemma-7b was gated too, under its own terms. The names on that list running under free licenses were DeepSeek under MIT, Mistral under Apache 2.0, and OpenAI’s gpt-oss under Apache 2.0. Qwen, which dominates the download slice, is Apache 2.0 as well.
| Model family (measured repo) | License | Gating |
|---|---|---|
| Meta Llama (2 / 3 / 3.1 variants) | custom community license | approval required |
| google/gemma-7b | Gemma terms | approval required |
| DeepSeek | MIT | none |
| Mistral 7B | Apache 2.0 | none |
| OpenAI gpt-oss | Apache 2.0 | none |
Gating is not a cosmetic restriction. An Apache model comes down with one line of code, while a gated model makes you clear an approval step before you can touch the weights.
My read
The data draws an inverse relationship between branding and actual rights. Meta and Google, the two leaning hardest on the word “open,” attach the most closed distribution terms. The labs that widened real adoption, including OpenAI’s own gpt-oss, did it under permissive licenses. That reading rests on two observations.
- Among the top flagships by likes, every Meta Llama variant and google/gemma-7b combined gating with a custom license (measured).
- In the same set, DeepSeek, Mistral and OpenAI’s gpt-oss ran under Apache or MIT with no gating (measured).
An Open Weight License Is Not Open Source
The terms deserve a clean split. Open weight means you can download the finished weights and run them. Open source means you also get the training code and enough data information to reproduce and audit the model. The Open Source Initiative published its Open Source AI Definition, OSAID 1.0, in October 2024, and it requires exactly that training data information plus a license with no field-of-use restrictions.
By that bar, Llama, Gemma and Qwen all land as open weight, not open source. The OSI has said in plain terms that Meta’s Llama license is not open source. The Llama community license carries a clause requiring any company above 700 million monthly active users in the prior month to request a separate grant from Meta, granted at Meta’s sole discretion.
So is Apache-licensed Qwen fully open source? No. The license is free, but the training data stays closed, so under OSAID it remains open weight. The practical world therefore splits into three layers. Fully permissive weights you can ship, custom or gated weights with conditions attached, and a gray zone with no license at all.
The 14 Percent Gray Zone
That third layer is the dangerous one. Fourteen of the top 100 carried no license tag whatsoever. A missing license is closer to “no rights granted” than to “do what you like,” which leaves the legal basis for commercial use unclear.
An honest data caveat belongs here too. The download charts mix in CI test repositories like trl-internal-testing/tiny-Qwen2ForCausalLM and hmellor/tiny-random-LlamaForCausalLM. Those are more likely automation artifacts hit on repeat than real-world usage. So the numbers here should be read as a license distribution trend, not as absolute rankings. Positions and counts will keep shifting after the snapshot date.
What To Actually Pick
The practical takeaway is simple. For a model going into a commercial product, Apache 2.0 and MIT are the cleanest choices. Gating, custom licenses and missing tags are each a reason to loop in legal.
Read the license text and the gating flag, not the word “open.”
The license field and access terms on a model card take a few minutes to check. Those minutes are what keep you clear of a 700 million user clause or a gray-zone repository.
The Most Downloaded AI Model Isn’t a Chatbot
References
- Own measurement (Hugging Face public API, 2026-07-25 KST). License and gating tally across the top 100 by downloads and the top text-generation models by likes. Method and raw output are in the methodology paragraph.
- Open Source Initiative (2025-02-18). Why Meta’s Llama license is not open source.
- Llama 3.1 Community License (2024-07-23). The 700 million MAU clause, acceptable use policy, and output restriction.
- meta-llama/Llama-3.1-8B-Instruct model card. Gated access and license tag confirmation.
- RedMonk (2026-05-12). Hugging Face license distribution analysis and the open-washing debate (context).
Stay Ahead of the Curve
Get the latest insights on AI, tech, and markets — delivered in both Korean and English.
Jugger | Backend Developer
A backend engineer working in enterprise solutions software since June 2021. The hands-on perspective from system design, API implementation, and production operations directly informs this site’s
AI, IT, and cloud infrastructure coverage.
For topics outside direct professional experience—blockchain and macroeconomics—analysis is built on primary-source citations such as SEC filings, whitepapers, institutional research, and government
statistics. Speculation is clearly separated from verified fact, and unverified claims are not published.
AI tools assist with drafting, but every article is reviewed, revised, and verified before publication. Errors are corrected with visible notes when discovered.
Contact: jjunggulee@gmail.com