Mistral Medium 3.5

Mistral Medium 3.5 (April 29, 2026) is Mistral's first open-weight mid flagship: a 128B dense model that merges the former Medium 3.1 (instruct), Magistral (reasoning) and Devstral 2 (coding) lines into one checkpoint with a per-request reasoning-effort knob. It scores 77.6% on SWE-bench Verified, takes text and image input over a 256K context window, and its weights are downloadable from Hugging Face under a Modified MIT license — a break from the API-only Medium 3.

The license deserves a careful read: 'Modified MIT' permits commercial use but carves out high-revenue companies (reportedly a $20M/month revenue bar), so it is neither fully open nor closed. As a dense 128B model it also needs serious hardware (~134GB in FP8, ~250GB in BF16) — realistically 4-8 GPUs. For teams that want one predictable, self-hostable model for instruct, reasoning and coding workloads without MoE memory quirks, it is a compelling pick.