Alpha
Contact
Implement · Mistral

Can you run it?

Ownership levelLimitednone·limited·partial·substantial·fullAnalytical input C ยท 63.2/100

This page is a projection of the one entry record, the Reliability and Data control factors that Implement covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Install & run

Download the safetensors from the verified mistralai organisation on Hugging Face and serve with vLLM, transformers, llama.cpp, or Ollama, for research / non-production use only. Pin the exact revision, verify checksums, and confirm the per-model licence tier (MRL vs MNPL).

Hardware & VRAM requirements

A wide range: Ministral 8B and Codestral 22B run on one or two GPUs; Mistral Large 2 (123B) and Pixtral Large (124B) need serious multi-GPU infrastructure or aggressive quantization. Any production deployment is licence-barred without a commercial agreement.

Serving stacks

vLLM, transformers, llama.cpp, and Ollama have strong mechanics, but production serving is licence-barred without a commercial agreement.

Safe-deployment controls & Deployment Ceiling

Deployment Ceiling: T2 (conditional). Do not deploy in production or commercially without a negotiated Mistral licence; the MRL/MNPL bar it. Supply your own input/output guardrails and a guard/classifier model (historically lighter safety tuning, no first-party guard), add prompt-injection defences, and treat images as untrusted input for Pixtral Large.

Available quantizations

Community GGUF / AWQ / FP8 builds circulate, but their use stays bound by the MRL/MNPL non-commercial terms. A quant confers no commercial rights.

Fine-tuning & adaptation

Fine-tuning for research is permitted, but derivatives carry the non-commercial bar: commercial/production use of a derivative needs a negotiated Mistral licence. Training data and code are closed, so there is no from-scratch reproduction.

API / OpenAI-compatible integration

Serve behind vLLM's OpenAI-compatible endpoint for research. The hosted API versions of Mistral Large 2.x, Pixtral Large and Ministral 8B are retired (Mistral docs), so the licensed commercial path is a negotiated commercial licence from Mistral or a current hosted model, not these checkpoints on the API.

How this scores

The ownership factors this domain covers, drawn from the one entry record.

3

ReliabilityIs it reliable and good enough for the job?

Strong

Performance 4, operational 4 and safety 3 from an accountable EU provider: capable models with strong serving. All are at or above 3 and two are at 4, so reliability is strong. The caveat is the absence of a first-party guard model.

How this scores (AOI sub-dimensions)
Operational4/5how practical it is to run, serve and maintain in productionServing mechanics are strong (vLLM, transformers, llama.cpp, Ollama) and community quants exist.
Safety3/5whether misuse risks are evaluated and guardrails are providedSafety-tuned instruct/multimodal models with documented behaviour meet the score-3 anchor.
4

Doesn't extract your dataDoes running it keep your knowledge and data yours?

Strong

Self-hosted for research, the weights run entirely on your own infrastructure with no telemetry or clawback, so your data stays yours - the constraint is on permitted use, not data residency.

How this scores
Not a scored AOI dimension. For a self-hosted model, data-control is a structural property of running the weights yourself, strong by default unless the model phones home or the licence claws back rights. For a hosted API this factor is the retention + train-on-inputs + residency read, scored from the binding terms.
What this means for adoptionYou have only limited ownership of these Mistral models. Self-hosting keeps your data yours and the models are capable and well-supported (data-control and reliability strong). The MRL/MNPL bar commercial and production use without a separately negotiated licence, so use-and-modify is weak and ownership is limited. For anything beyond research or evaluation, negotiate a commercial licence with Mistral or use the Apache-2.0 Mistral models (the `mistral` entry), which carry none of these restrictions (that entry is rated partial overall only because the family is licence-split). Mistral Large 3 and Ministral 3 are Apache-2.0 successors; Mistral Large 4 uses a different custom licence.

Sources

The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.

Licenceread2026-11-19
Mistral licence texts.
Model cardread2026-11-19
mistralai Hugging Face org (verified), model cards and API: Ministral-8B-Instruct-2410, Mistral-Large-Instruct-2407 and 2411 and Pixtral-Large-Instruct-2411 carry license_name mrl; Codestral-22B-v0.1 carries license_name mnpl.
Documentationread2026-11-19
Mistral docs Models Overview: no model carries an MRL or MNPL label.
Third-party analysisunverified2026-08-03
On public leaderboards Mistral Large 2 and Pixtral Large are strong open-weight models, and Ministral 8B is strong for its size.
Third-party analysisunverified2026-08-03
Mistral has historically shipped lighter safety tuning than some peers and no first-party guard model; independent behavioural analysis is limited.
Third-party analysisread2026-11-19
Mistral AI is EU-domiciled (France) and is listed among the 22 signatories of the GPAI Code of Practice on the European Commission page (last updated 7 October 2026).