Table of Contents
For commercial deployments, open-weight model licensing matters. Some are fully permissive (Apache 2.0); others have specific restrictions (Llama 3 acceptable use, Gemma terms, Mistral Non-Production License). Read before committing to a model in production.
Most permissive: Apache 2.0 (Mistral 7B / Mixtral / Qwen 2.5 / Yi / DeepSeek). Permissive with specific terms: Llama 3 Community Licence (commercial-friendly with caveats above 700M MAU), Gemma terms. Source-available: Mistral Small 3 / Codestral (Mistral AI Research / Non-Production Licence — verify your use case).
Licensing
| Model family | Licence | Commercial use |
|---|---|---|
| Mistral 7B / Mixtral 8x7B | Apache 2.0 | Yes, freely |
| Mistral Small 3 / Codestral | Mistral AI Non-Production / Research | Limited — verify with Mistral |
| Llama 3.1 / 3.3 | Llama 3 Community Licence | Yes (with attribution + AUP, < 700M MAU) |
| Qwen 2.5 (all sizes) | Apache 2.0 (most variants) | Yes, freely |
| Yi 1.5 | Apache 2.0 | Yes, freely |
| Gemma 2 | Gemma terms | Yes (with prohibited-use policy) |
| DeepSeek V3 / R1-Distill | MIT / Apache 2.0 (variant-dependent) | Yes |
| Phi-3 / Phi-4 | MIT | Yes, freely |
Commercial use
Three notable cases:
- Llama 3 700M MAU clause: if your service has > 700M MAU, you need a separate licence from Meta. Almost no one hits this.
- Codestral / Mistral Small 3 licence: Mistral's "Non-Production" / "Research" tier. Verify your specific use case; commercial use may require ongoing licensing.
- Gemma prohibited use: Google's prohibited use policy excludes specific harmful uses. Standard for permissive models in 2026.
For maximum permissiveness with no caveats: Mistral 7B / Mixtral 8x7B / Qwen 2.5 / Phi family.
Verdict
Read the licence of every model before production deployment. For most workloads, Apache 2.0 / MIT models (Mistral 7B, Qwen 2.5, Phi) provide maximum freedom. Llama 3 family is fine for almost everyone (700M MAU is rarely hit). Mistral's commercial-tier products (Codestral, Mistral Small 3) require licence diligence.
Bottom line
Apache 2.0 / MIT for max freedom. Verify Mistral's specific licences. See 2026 model rankings.
