Mistral Medium 3.5 — 128B dense, 256K context, 77.6% on SWE-Bench Verified
Mistral's April 29 release ships under a modified MIT license, with 77.6% on SWE-Bench Verified — positioning the model ahead of Devstral 2 and Qwen 3.5 397B A17B at a fraction of the active-parameter budget.
Architecturally, Medium 3.5 is a dense (not MoE) 128B model with a 256K token context window. The "modified MIT" wording suggests added clauses around commercial restrictions or competing-service prohibitions — Mistral's licensing has historically threaded the open/proprietary line carefully, and the modified-MIT pattern is consistent with that posture.
The SWE-Bench Verified number is the headline — 77.6% puts Medium 3.5 in the same band as Claude Code's coding-tier performance, which had been considered out of reach for dense open-weights models below the 400B class. If the number replicates at scale on third-party evals, this is a meaningful narrowing of the open/closed gap for autonomous coding.