مسٹرل میڈیم 3.5

Mistral's frontier-class multimodal model, optimised for agentic and coding use. It is the lab's first flagship merged model -- a dense 128B set of weights with a 256K context that folds the previous Medium, Magistral reasoning and Devstral coding lines into one, handling instruction following, reasoning and coding together. It accepts text and image input and exposes a per-request reasoning toggle, so the same weights serve a quick reply or a long agentic task.

Medium 3.5 did not arrive as a standalone model launch. Mistral announced it as the thing that made a product possible: coding agents that leave the laptop and run in the cloud, in parallel, for as long as a job takes. The lab is direct about the causality - it calls Medium 3.5 the model that made asynchronous cloud agents practical to ship. On the day it landed it became the default in Le Chat, replaced the lab's dedicated coding model inside its Vibe CLI, and became the engine behind a new multi-step Work mode for research, analysis and cross-tool actions.

That shapes what the model was built to do. Mistral trained it for long-horizon runs, for calling several tools reliably in one turn, and for producing structured output that downstream code can consume without a parser full of special cases. Reasoning effort is set per request, so the same weights answer a one-line chat message or grind through an hour-long agentic task. The vision encoder was trained from scratch to cope with variable image sizes and aspect ratios rather than squaring everything to a fixed input.

The comparisons Mistral published are all run at maximum reasoning settings and are mostly about size. It reports 77.6% on SWE-bench Verified, ahead of its own previous coding models at 72.2% and 68.0%, and it reports leading the far larger open models it compared against on the same set. On the multi-domain agent suites it reports 91.4 on the telecom split, 76.1 on retail and 72.0 on airline, with 69.0 on the AllenAI instruction-following benchmark IFBench, 95.8 on Collie, and 86.3 on AIME 2025 averaged over sixteen runs.

Where it stops is a matter of scale and of domain. On Mistral's own chart the banking split of the agent suite sits at 13.4 - far below the same model's telecom and retail results and the weakest column it published - and its web-browsing number is reported with context management and a discard-everything strategy at 100k tokens rather than as a plain run. This is a 128 billion parameter dense model whose selling point is that it can be self-hosted on as few as four GPUs and still hold its own; Mistral makes the efficiency argument, not a frontier-capability one.