Not legal advice. Licenses change, so read the current text before you build a business on one.
Removing a model’s refusals is a modification, and a modified model is a derivative. Derivatives inherit the base model’s license. An abliterated Llama is still governed by Llama’s license and acceptable-use policy, and an abliterated Qwen by Qwen’s.
This matters if you’re choosing a provider: a host selling a derivative it isn’t licensed to sell is a host that could disappear overnight. It also matters if you’re thinking of hosting one yourself.
The cheat sheet
| Base model family | License | Commercial hosting of an abliterated derivative |
|---|---|---|
| Qwen3 / Qwen3.5 | Apache-2.0 | ✅ Yes, keep the license notice |
| Qwen3.8 | Its own Qwen3.8-Max License | Not Apache. Read the terms before hosting |
| Gemma 4 | Apache-2.0 | ✅ Yes |
| Mistral Small 3.x / Large 3 | Apache-2.0 | ✅ Yes |
| gpt-oss | Apache-2.0 + short usage policy | ✅ Yes, follow applicable law |
| GLM-4.6, DeepSeek-V3.2 | MIT | ✅ Yes |
| Qwen2.5-72B | Qwen License | ⚠️ Yes. A model you fine-tune and distribute must say “Built with Qwen”; a separate license is needed above 100M monthly users |
| Llama 3.3 | Llama 3.3 Community License | ⚠️ Yes, with conditions (below) |
| Gemma 1–3 | Gemma Terms of Use | ⚠️ Yes, with conditions (below) |
| Qwen2.5 3B models | Qwen Research License | ❌ Non-commercial only (other Qwen2.5 sizes up to 32B are Apache-2.0) |
| Mistral-Large-2411 and others | Mistral Research License | ❌ No commercial hosting; SaaS explicitly excluded |
| Kimi-K2, MiniMax-M2 | Modified MIT | ✅ Yes, but very large services (100M+ monthly users or $20–30M revenue) must show the model’s name in the UI |
The Llama conditions
The Llama 3.3 Community License allows derivatives and commercial use, but:
- You must display “Built with Llama”.
- A derivative model you distribute or make available must have a name that begins with “Llama”.
- You must include the license.
- The Acceptable Use Policy is incorporated into the license. You agree not to use Llama, or “allow others to use” it, for things like weapons development or fraud, and to disclose known dangers of your AI system to end users.
- Companies with over 700M monthly users need a separate license.
Abliteration doesn’t remove the acceptable-use policy. The policy never mentions derivatives explicitly, but a host serving an abliterated Llama is still using Llama under this license, so in our reading it’s still bound by it. That creates an obvious tension for a product marketed as “no rules.”
The Gemma conditions
The Gemma Terms for Gemma 1 through 3 treat hosted API access as distribution. You must pass the Prohibited Use Policy on to your users as an enforceable term, and Google reserves the right to restrict usage “remotely or otherwise”. Gemma 4 switched to Apache-2.0, which removes those strings, and that’s one reason Gemma 4 derivatives are spreading fast across uncensored hosts.
The trend: open licenses are winning
The newest strong open models (Qwen3 and 3.5, Gemma 4, Mistral’s 2025–26 releases, gpt-oss, GLM, DeepSeek) are almost all Apache-2.0 or MIT. For uncensored hosting, that’s the whole game: a permissive base means an abliterated derivative is as legally clean as the original.
What we track
Every model in the index carries its base license where we can identify it. Our methodology flags providers charging for derivatives of research-only models. A license problem isn’t a score penalty, but you should know before you build on that endpoint.