Apache 2.0 open weights from 3B to large, which is the difference that decides it for teams whose legal review stalled on Llama's conditions. Read the licence per model though: the catalogue mixes Apache, modified MIT, commercial and one that prohibits commercial use.
Apache 2.0, which is the whole argument
Meta Llama is the cautionary case: open weights under a licence that is not open source, with conditions and a user threshold to read before you build.
Mistral's open line is Apache 2.0. No acceptable-use rider, no scale threshold, no attribution requirement beyond the licence itself. For a team whose legal review stalled on Llama's terms, that difference decides it before any benchmark does.
The Apache-licensed models span sizes rather than being one release: Mistral Small 4, Mistral Large 3, and Ministral 3 at 14B, 8B and 3B.
More than a chat model
The second reason to look is breadth. The family covers jobs that usually mean adding another vendor:
| Need | Model |
|---|---|
| Transcription | Voxtral Small, Voxtral Mini Transcribe |
| Speech synthesis | Voxtral TTS |
| Document OCR | OCR 4.1 and earlier |
| Code generation | Codestral |
| Code embeddings | Codestral Embed |
| Semantic embeddings | Mistral Embed |
| Safety filtering | Mistral Moderation 2, Shieldstral 1.0 |
Shieldstral and the Voxtral transcription models are Apache 2.0, so the safety and speech pieces can run on your own hardware too. That is unusual, and it matters for the case security and privacy concerns describes, where audio of customer calls cannot go to a third party.
Read the licence per model, not per vendor
The catalogue mixes licences, and assuming "Mistral is Apache" will eventually be wrong.
Mistral Medium 3.5 is a modified MIT licence. Codestral, the OCR models, Mistral Embed and Mistral Moderation 2 are Premier, which is commercial and API-only. Voxtral TTS is CC BY-NC 4.0, which prohibits commercial use.
That last one is the trap. A non-commercial licence on a text-to-speech model is easy to miss in a prototype and expensive to discover after launch. Models on Hugging Face makes the same point about reading the licence first, and this catalogue is a good demonstration of why per-model rather than per-vendor.
The platform also serves models Mistral did not train, including Z.ai GLM releases with 1M-token context. Useful, and another reason the licence question is per model.
Where it fits
Reach for Mistral when the licence is what is blocking you, when you want a size range from 3B to large under one family and one set of terms, or when you need transcription, OCR or moderation without adding vendors.
The small end deserves attention. Ministral 3 at 3B and 8B is sized for a laptop or a modest server, which puts it beside Gemma and makes Ollama or LM Studio the natural way to try it.
As with any open-weight family, running the weights is a commitment rather than a saving, and self-hosted models covers what you take on.
Evaluating it
The same discipline as any model choice. Choosing the right model argues that twenty cases from your own workload beat a leaderboard, and it applies harder to a family this wide: a benchmark on Mistral Large tells you nothing about whether Ministral 3B is sufficient for your classification step, which is the question that decides your bill.
Further reading
- Mistral, Models overview: the catalogue, with a licence against each model.
- Meta Llama: open weights under a licence that is not open source.
- Self-hosted models: what running the weights involves.
- Choosing the right model: the comparison to run.
Knowledge check
Question 1 of 3
Sign in to save your progress and pick up where you left off.