Can you own it?

Intended & out-of-scope use

Mistral's open-weight family is a general-purpose set of assistant, coding and multilingual text models, spanning the dense Mistral 7B, the Mixtral 8x7B / 8x22B mixture-of-experts models, the 128K-context Mistral Nemo (12B) and the 24B Mistral Small 3.x, plus specialist variants (Codestral for code, Mathstral, Devstral, Pixtral for vision). Its realistic differentiator is being the leading European open-weight family: EU-domiciled, strong multilingual coverage, and the cleanest GPAI regulatory posture in this class.

For adoption, deploy the instruct (safety-tuned) variant, not the base checkpoint, and - uniquely for this family - confirm the specific model's licence before any commercial deployment (the Apache flagship models are unconditional; the MRL research-only models are not). Mistral is not positioned as a safety-critical model; high-stakes, autonomous or regulated decision-making is out of scope without the external control stack described in the Implement domain.

Known limitations, bias & failure modes

• Licence split is the headline risk. A team that assumes "Mistral = Apache" can trip over the research-only MRL models (Ministral 8B, Mistral Large, Pixtral Large) and the originally non-production MNPL Codestral. Commercial use of those requires a negotiated licence.
• Lighter safety tuning. Instruct alignment withstands casual jailbreaks, but coverage is historically lighter than the largest US labs, and the companion classifier (the Moderation API) is a hosted service rather than a broadly-shipped open guard-weight.
• Closed training data and code. Unlike a fully-open family, you cannot inspect or reproduce the training pipeline, so bias sources must be reasoned about from an opaque artifact.
• Standard hallucination and prompt-injection profile for the class; treat retrieved/tool content as untrusted.

The offsetting advantage is regulatory: an EU-domiciled, Code-of-Practice-signatory provider with a genuine OSI licence on its flagship models.

Openness tier & components

Mistral earns open_weights honestly. The flagship general-purpose models ship downloadable Apache-2.0 weights - an OSI-approved permissive licence, not a click-through community gate - and documentation is good. But training data and training code are closed (no reproduction path, unlike a fully-open family), evaluation is only partial, and the licence component is split: a meaningful subset of the family is research-only MRL. That mix is why Dimension 1 scores 4, above a gated-open family but below the fully-open exemplars.

License terms & permitted use

There is no single licence - read the model card:

• Apache-2.0 (OSI-approved, unconditional commercial use): Mistral 7B, Mixtral 8x7B / 8x22B, Mistral Nemo, Mistral Small 3.x, Magistral/Devstral Small. You may use, modify, redistribute and commercialize derivatives subject only to standard attribution/notice terms. Mistral has publicly recommitted to Apache-2.0 for general-purpose models with Mistral Small 3.
• Mistral Research Licence (MRL-0.1), research-only: Ministral 8B, Mistral Large, Pixtral Large. Any commercial deployment requires a separately negotiated licence from Mistral.
• MNPL (non-production): originally Codestral.

This split is a load-bearing input to both the openness (4) and legal (4) scores - the Apache core is genuinely clean, but the family is not uniformly open.

Supply-chain & provenance

Weights are distributed from the verified mistralai org on Hugging Face in safetensors (no pickle requirement) with per-file checksums. The checkpoint trust checklist scores about 6/8; the two missing controls are cryptographic weight signing (Sigstore/model-signing) and SLSA build attestation, which is why provenance is a 4, not a 5. No malicious-checkpoint incident is on record for the canonical org. Community GGUF/AWQ/FP8 quants and hosted cloud endpoints (Bedrock, Azure AI, Vertex, Together, OpenRouter) are separate artifacts - pin the exact mistralai revision and verify checksums on download.

EU AI Act posture

This is Mistral's standout. The open-weight family is a GPAI model set that sits well under the 10²⁵-FLOPs systemic-risk threshold (the Apache releases are ≤24B dense or Mixtral MoE; only the 123B Mistral Large 2 approaches it, and that is a research-only MRL model, not an Apache release), so no Article 55 regime applies. The flagship Apache-2.0 models plausibly qualify for the Article 53 open-source exemption from the Annex XI/XII technical-documentation duties; the research-only MRL models, carrying a field-of-use restriction, do not. Two further facts make Mistral the most EU-AI-Act-friendly non-US family here: it is EU-domiciled (Paris), and it was an early signatory of the EU GPAI Code of Practice - with no EU field-of-use ban of the kind that affects some competitors. The surviving copyright-policy and training-content-summary obligations are only partially met, because Mistral does not publish its training corpus. A fine-tuner placing a derivative on the EU market may become a provider for that derivative.

Benchmarks & evaluation

OneHill did not run its own benchmarks this session; the picture below is aggregated from independent public leaderboards and third-party coverage (ev-leaderboard, ev-wiki). Mistral models are competitive-to-strong within each size class: Mistral 7B and Mixtral were class-leading at release, and Mistral Small 3.x remains strong for a 24B. Mistral is consistently judged the leading European open-weight family. Treat published scores as third-party/in-class, not as an OneHill-verified claim of category leadership - performance is capped at 4 for exactly that reason.