Mistral

TAS Score: S3/3 — D4/5 — A3 / T2 (D4 not D5: the hosted La Plateforme API collects usage data and requires an account. The self-hosted open-weight variant scored here carries none of this.)

Family: Mistral (Apache open) · Mistral Medium (Modified MIT) This card scores the Apache-2.0 open-weight models. The Modified-MIT tier (Mistral Medium 3.5, Devstral 2) carries a non-OSI revenue clause and is a separate object — see the family link.

⚠️ Two-mode tool. Self-hosted open weights: A3/T2. Hosted La Plateforme API: A1/T1. This card scores the self-hosted open-weight variant.

Brief Description

Open-weight model family from Mistral AI (Paris) under the Apache-2.0 license — Mistral 7B, Mixtral 8x7B/8x22B, Mistral Nemo, Pixtral 12B, Mathstral, Ministral 3, Mistral Small / Small 4, Magistral Small, Devstral Small, and the flagship Mistral Large 3 (41B active / 675B total MoE, released Apache in December 2025). All fully downloadable and self-hostable; sizes span a laptop-friendly 7B up to a multi-GPU 675B MoE, with EU data-residency positioning.

Scope note: a few Mistral models sit under more restrictive licenses — Modified MIT (Medium 3.5, Devstral 2), the Mistral Research License (Ministral 8B, Mistral Large 2, Pixtral Large), and the non-production MNPL (classic Codestral). Those are not Apache; the Modified-MIT tier is scored on the family card, and MRL/MNPL models are noted there.

Architectural Role

Compute/inference layer: general-purpose open-weight LLMs run on your own hardware, with an OpenAI-compatible API for drop-in self-hosting. Local alternative to cloud chat, coding, and reasoning APIs across the full size range.

Technical Autonomy

  • Works without internet (after model download)
  • Stores data locally
  • Does not require external accounts (self-hosted)
  • Allows data export — Apache-2.0, standard HuggingFace format
  • Hosted La Plateforme API requires an account and sends data to Mistral servers

Philosophical Assessment (whose.world criteria)

Criterion Status Comments
Pause ✅ Stop inference. Model weights stay on disk.
Exit ✅ Standard model format. Apache-2.0 — no restrictions on use or redistribution.
Recoverability ✅ Re-download from HuggingFace or restore from backup.
Visibility ✅ Apache-2.0 across the core lineup. Open weights. Training data not published.
External Dependencies ⚠️ Self-hosted: none. Hosted API: Mistral infrastructure (EU-based).

Configuration (Minimal) 

# Consumer hardware — via Ollama
ollama run mistral-nemo      # 12B, Apache-2.0
ollama run mixtral           # 8x7B MoE, Apache-2.0

# Server — via vLLM (Apache-licensed checkpoint)
vllm serve mistralai/Mixtral-8x22B-Instruct-v0.1 --tensor-parallel-size 4

Alternatives

Alternative Autonomy Notes
Qwen3 A3 / T2 Apache-2.0 open-weight family.
DeepSeek-R1 A3 / T2 Open-weight reasoning, MIT.
Mistral Medium A3 / T1 Same lab — Modified-MIT revenue-clause tier.

Trajectory

Direction: opening

Mistral is the clearest permissive-open bet among major labs — Large 3, Small 4, Ministral, Mixtral, and Nemo all ship under Apache-2.0, with EU data residency. The caveats are a Modified-MIT tier (Medium 3.5, Devstral 2) with a revenue clause and a few research / non-production licenses; the core lineup stays Apache.

Signal assessment:

Signal Status Evidence
License ✅ Core lineup Apache-2.0; Mistral Large 3 released Apache (Dec 2025).
Feature gating ➖ Some flagship / coding models under restrictive licenses, but the Apache core is broad.
Self-hosting ✅ Standard formats. Ollama, vLLM, SGLang, llama.cpp; OpenAI-compatible API.
Governance ✅ Independent EU company. Active, frequent open releases.

Signal key: ✅ opening · ➖ neutral · ⚠️ closing

Sources

This site uses Just the Docs, a documentation theme for Jekyll.