Mistral's first merged flagship: 128B dense, single set of weights folding instruction following, reasoning, and coding into one model. Replaces the prior Magistral / Devstral 2 split. 256K context. Configurable reasoning-effort setting. Trained-from-scratch vision encoder handling variable image sizes and aspect ratios. Released April 29, 2026 under a modified MIT license, open weights, runs on 4 GPUs.

AA Intelligence Index: 39 — Mistral's highest to date and well above the open-weight median of 15. 77.6% on SWE-Bench Verified. Powers the Vibe remote agents (async cloud coding from CLI or Le Chat) and Le Chat's new Work mode.

Model Details

Architecture DENSE
Parameters 128B
Context window 256,000
AA Intelligence 39

Benchmark Scores

Benchmark Score Mode
SWE-Bench Verified 77.6%
frontierreasoningcodingmultimodalopen-weight

Related