Mistral Medium 3.5 Retires Three Separate Models Into One 128B Open Weight

Mistral just did something the other frontier labs have been doing quietly: they collapsed their entire model lineup into a single set of weights. Mistral Medium 3.5 is a dense 128B parameter model with a 256K context window that replaces three previously separate models in the stack, and it ships under a Modified MIT license so you can actually run it yourself.

Shipping Medium 3.5 also quietly retired Magistral (reasoning), Devstral 2 (coding), and Medium 3.1 (chat). Three product lines folded into one 128B dense model with a reasoning toggle. That consolidation is worth more attention than any single benchmark number.

One model, one toggle, two personalities

The core trick is a reasoning_effort parameter you set per request. Set it to none for fast instruction-following, or high to unlock chain-of-thought traces wrapped in [THINK] tags for hard problems. Same weights, different behavior, chosen at call time.

Compare that to what the industry has been doing. There's a "fast" or "instant" checkpoint trained primarily for chat-style instruction following, and a separate "thinking" checkpoint heavily post-trained on chain-of-thought traces. When you toggle "thinking" in ChatGPT or Claude, you route your prompt to a different set of weights. Medium 3.5 collapses this entirely: one weight file, two modes, no router in front.

The benchmark story

Medium 3.5 supersedes all of Mistral's previous coding models across benchmarks, scoring 91.4% on τ³-Telecom and 77.6% on SWE-Bench Verified. The τ³ (tau-cubed) suite measures multi-turn tool use across telecom, airlines, retail, and banking, so it works as a proxy for agentic reliability rather than a static QA score. On the strength of those numbers, Medium 3.5 also takes over from Devstral 2 inside Mistral's Vibe CLI.