License summary

A side-by-side cheat sheet covering the licenses for each model in the catalog. This is not legal advice. For any commercial deployment, read the actual license text and consult counsel; the table below is a quick orientation only.

Model	License	Commercial use	Redistribution	Notable clauses
Llama 3.3 70B	Llama 3.3 Community	Allowed*	Allowed (with attribution)	* Separate agreement required if your service has > 700M MAU at release
DeepSeek-V3	MIT	Allowed	Allowed	No restrictions
DeepSeek-R1	MIT	Allowed	Allowed	Distilled variants inherit base license (Qwen / Llama)
Qwen2.5-72B	Qwen License	Allowed*	Allowed (with attribution)	* Service-scale clause at 100M MAU; smaller Qwen2.5 variants are Apache 2.0
Mixtral 8x22B	Apache 2.0	Allowed	Allowed	No restrictions
Mistral 7B v0.3	Apache 2.0	Allowed	Allowed	No restrictions
Gemma 2 27B	Gemma Terms of Use	Allowed	Allowed (terms travel)	Acceptable-use policy must accompany derivatives
Phi-4	MIT	Allowed	Allowed	No restrictions
Command R+ (08-2024)	CC-BY-NC 4.0	Not allowed	Non-commercial only	Commercial use requires separate Cohere license
Yi-1.5 34B Chat	Apache 2.0	Allowed	Allowed	No restrictions
GLM-4 9B Chat	GLM-4 Model License	Allowed	Allowed (with notice)	Notification required for large-scale deployment
Falcon 3 10B	Falcon LLM License 2.0	Allowed	Allowed	Apache-derived with acceptable-use clause
OLMo 2 13B	Apache 2.0	Allowed	Allowed	Training data licensed per-source — see Dolmino corpus docs
SmolLM2 1.7B	Apache 2.0	Allowed	Allowed	No restrictions
Granite 3.1 8B	Apache 2.0	Allowed	Allowed	IBM provides indemnification when used via watsonx.ai

If you only have time to read one thing

The safest commercial-use bucket is Apache 2.0 and MIT: nine of the fifteen catalogued models fall there. Llama, Gemma, Qwen, and GLM are permissive enough for most commercial work but have specific clauses worth a five-minute read. Command R+ is non-commercial — do not deploy it in production without a separate Cohere license.