Mistral Models

Overview of Mistral’s models with fast inference and strong coding performance.

Latest: Mistral Medium 3.5

Mistral Medium 3.5 (Apr 2026) is Mistral's flagship merged model, unifying instruction-following, reasoning, and coding into one 128B dense model with a 256k context window. It's self-hostable on as few as four GPUs and released under a modified MIT license.

Strengths

  • 256k context window
  • Unified chat/reasoning/coding weights
  • Self-hostable, EU-based

Best For

  • Production assistants
  • Coding agents (replaces Devstral 2)
  • Privacy‑sensitive workloads

Model Lineup

Mistral Medium 3.5

Flagship merged model for chat, reasoning, and coding.

Mistral Large 3

Open-weight (Apache 2.0) MoE flagship, strong multilingual/vision performance.

Ministral 3 (3B/8B/14B)

Edge-optimized open-weight family with vision support.

Practical Guidance

  • Use the built-in reasoning_effort parameter to trade off speed vs. depth per request.
  • Mistral Medium 3.5 now replaces Codestral/Devstral 2 in Mistral's own coding agent (Vibe CLI).
  • Prefer EU hosting options for privacy requirements.