Tier 3 Practical open-weight
What the publisher claims: open models, open-weight
The one-liner: The permissive half of a portfolio whose openness depends on the exact release.
The six questions
Downloadable weights for the whole permissive line, from 3B Ministral models to the 675B Large 3.
Plain Apache 2.0 with no revenue threshold, usage policy, or attribution condition on these releases.
Self-hostable on chosen hardware with documented vLLM, TensorRT-LLM, TGI, and Ollama paths; Small 4 runs on a single DGX B200 while Large 3 needs a multi-GPU node.
Model cards and technical materials vary by release and are strong on architecture and benchmarks, but training corpora and recipes are not consistently public.
Moderate to limited: published materials describe the models rather than the process that built them.
Mistral maintains a multi-repo public GitHub org — mistral-common, mistral-inference, cookbook — with published contribution guides where outside pull requests are routinely merged, and the Apache-2.0 line ships code alongside weights, though Mistral alone sets the roadmap and releases.
License
Apache 2.0 on Large 3, Small 4, the Ministral 3 family, Devstral Small 2, Magistral Small, and the Voxtral and Shieldstral models OSI-approved — license source ↗
- Other releases in the portfolio carry different terms; check the exact model and license
What is actually released
- weights (Large 3 675B, Small 4 119B, Ministral 3 at 3B/8B/14B, Devstral Small 2, Magistral Small, Voxtral, Shieldstral)
- model cards
- technical materials (variable by release)
Why it matters
Large 3, Small 4, and the Ministral 3 family are the clean Apache 2.0 counterweight to the rest of Mistral's lineup, and they make the family-level 'open' claim impossible to read without checking the checkpoint.
Sources
- https://docs.mistral.ai/models
- https://vdf.ai/blog/mistral-on-premises-deployment
- https://myclaw.ai/blog/mistral-small-4-vs-mistral-large-3
- https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.