Mistral Medium

TAS Score: S3/3 — D4/5 — A3 / T1 (T1 not T2: the “Modified MIT” license adds a revenue clause — no rights if your company’s consolidated monthly revenue exceeds $20M — which is not OSI-approved. Weights are downloadable and self-hostable, so autonomy stays A3; only license transparency drops.) (D4 not D5: the hosted La Plateforme API collects usage data and requires an account.)

Family: Mistral (Apache open) · Mistral Medium (Modified MIT) This card scores the Modified-MIT tier — Mistral Medium 3.5 and Devstral 2. The Apache-2.0 core lineup is a separate object — see the family link.

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

Brief Description

Mistral’s Modified-MIT tier — Mistral Medium 3.5 (128B dense, 256K context, multimodal, released April 2026) and Devstral 2 (agentic coding). The weights are published on HuggingFace and fully self-hostable, but under a “Modified MIT” license that reads like MIT with one added condition: organizations whose consolidated monthly revenue exceeds $20M get no rights and must obtain a commercial license or use Mistral AI Studio. Because that revenue clause is not OSI-approved, transparency lands at T1 even though the weights are open.

This is the “looks open, isn’t OSI” case — the same shape as Meta’s Llama Community License (a scale-based usage restriction), and the reason DeepSeek-R1-Distill-Llama also sits at T1.

Architectural Role

Compute/inference layer: a frontier-class open-weight model (Medium 3.5) and an agentic coder (Devstral 2), self-hostable on multi-GPU hardware. Medium 3.5’s Q4 is roughly 75–80 GB — dual high-end GPUs, an H100/H200, or a 192 GB+ Mac.

Technical Autonomy

  • Works without internet (after model download)
  • Stores data locally
  • Does not require external accounts (self-hosted)
  • Allows data export — weights downloadable (Modified-MIT; restricted only for orgs with >$20M monthly revenue)
  • 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, downloadable. Modified-MIT restricts only >$20M-revenue orgs.
Recoverability ✅ Re-download from HuggingFace or restore from backup.
Visibility ⚠️ Open weights, but the Modified-MIT revenue clause is not OSI-approved → T1. Training data not published.
External Dependencies ⚠️ Self-hosted: none. Hosted API: Mistral infrastructure (EU-based).

Configuration (Minimal) 

# Devstral 2 (123B, agentic coding) via vLLM
vllm serve mistralai/Devstral-2-123B-Instruct-2512 --tensor-parallel-size 4

# Mistral Medium 3.5 — see the HuggingFace model card for the current checkpoint id

Alternatives

Alternative Autonomy Notes
Mistral (Apache) A3 / T2 Same lab — OSI-permissive core lineup, no revenue clause.
Qwen3 A3 / T2 Apache-2.0 open weights, no revenue clause.
DeepSeek-R1-Distill-Llama A3 / T1 Same “open weights, non-OSI license” shape.

Trajectory

Direction: mixed

Modified MIT keeps the weights open and self-hostable for nearly everyone, but the >$20M revenue clause is a deliberate monetization gate that OSI would not accept — a small step back from Mistral’s otherwise Apache-first stance. Whether future flagships stay Apache or drift into this tier is the signal to watch.

Signal assessment:

Signal Status Evidence
License ⚠️ Modified MIT — revenue clause, not OSI-approved.
Feature gating ⚠️ Merged flagship (Medium 3.5) placed in the restricted tier rather than Apache.
Self-hosting ✅ Weights downloadable; standard formats, vLLM / Ollama compatible.
Governance ✅ Independent EU company; license terms documented openly.

Signal key: ✅ opening · ➖ neutral · ⚠️ closing

Sources

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