Open Source LLM Licenses Explained: Llama vs Apache vs Gemma vs MIT

"Open source" doesn't mean what you think it means when it comes to LLMs. Of the 18 models in our directory, only 6 are actually open source by the OSI definition. The rest sit on a spectrum from "go ahead, ship anything" to "research only, talk to our lawyers." This is what you actually need to know.

True open source (OSI-approved) **Apache-2.0**: Used by Mixtral, Yi, OpenChat, OLMo 2, InternLM. The gold standard. You can build anything, ship anything, sell anything. Patent grant included. If you only want to ship products with zero license risk, filter for Apache-2.0.

MIT: Used by Phi-4. Even simpler than Apache — basically "do whatever, don't sue us." If you ever need to vendor a model into a closed-source product without disclosure obligations, MIT is the friendliest.

"Open weights" (not OSI but commercially friendly with caveats) **Llama Community License**: Used by Llama 4 family. Free to use, modify, redistribute, and even ship in commercial products — UNLESS your product has more than 700M monthly active users, in which case you must negotiate with Meta. There's also an acceptable-use policy banning specific applications (weapons, deception, etc.). For 99% of teams this is effectively MIT.

Gemma: Used by Gemma 3. Similar to Llama — commercial use allowed with use restrictions. The Gemma Prohibited Use Policy is broader than Llama's and you must agree to update terms over time. Big-co lawyers tend to push back.

Qwen License: Used by Qwen models. Commercial use allowed with the now-standard 100M MAU threshold. Mostly Llama-style.

DeepSeek License: Used by DeepSeek V4 and Coder V3. Permissive commercial use, no MAU cap, but with a use-case prohibition list (military, surveillance, etc.).

Tiered (revenue-based) **Falcon-LLM TII License**: Used by Falcon 3. Free below a revenue threshold (currently $1M/year), then you owe TII a fee. If you grow into the threshold mid-product-cycle this can be an unpleasant surprise — get sign-off from finance early.

Mistral Research License: Used by Mistral Large 3. Free for non-commercial research; commercial use needs a paid license from Mistral. Their permissive variants (Mixtral 8x22B Apache-2.0) cover most production needs.

Research only **Cohere Custom (research)**: Used by Command R+ 2 weights. You can download and study them, but you cannot ship them in a product. For commercial Command R+ 2 use, the Cohere API is the only path.

Open RAIL: Used by StarCoder 3. Permits commercial use but with a list of behavioral restrictions (no autonomous weapons, no defamation generation, etc.). Compliance burden is real for anything user-facing.

Picking by your situation - **Bootstrapped indie / small SaaS**: Apache-2.0 or MIT only. Why deal with anything else. - **VC-backed startup**: Llama, Gemma, Qwen are all fine. Most term sheets won't blink. - **Big-co lawyer-heavy environment**: Apache-2.0 only, full stop. Get pre-approval for anything else. - **Defense/government**: Read every license carefully. Open RAIL and DeepSeek explicitly prohibit defense use. - **Building a model service for resale**: Avoid anything with MAU caps unless you're sure you won't hit them.

The "100% open" question If you want everything — weights, training code, training data — there's basically only OLMo 2 (and to a lesser extent the Pythia family). Most "open source" LLMs publish weights only. Whether that matters to you depends on your reproducibility needs.