Mistral releases Medium 3.5 model

Language models are starting to split into two jobs. One job is chat. The other is work — long, messy, tool-using work that can keep going after you stop typing. Mistral’s April 29 release sits squarely in that second category. Medium 3.5 is not just another model drop. It is the engine Mistral is using to push its coding agent, its chat product, and a new “Work mode” toward longer autonomous runs. (mistral.ai)edium 3.5 in public preview on April 29, 2026, alongside two product hooks built around it: remote coding agents in Vibe and a new Work mode in Le Chat. Medium 3.5 also became the default model in both products, which tells you this is meant to be used immediately, not parked as a demo model off to the side. (mistral.ai) ### Why call it “Medium”(mistral.ai)iteral size claim. Medium 3.5 is a dense 128B model with a 256k context window. That is plenty large by normal standards. The point Mistral is making is that this class of model can hit flagship-style use cases without the deployment pain of the very biggest systems. The company says it can be self-hosted on as few as four GPUs. (mistral.ai)pposed to be good at? Mistral is pitching it as a merged model for instruction-following, reasoning, and coding. That matters because agent products break when you have to keep swapping between specialized models mid-task. Medium 3.5 is supposed to stay coherent across a long chain of actions — reading context, calling tools, writing structured output, then continuing without losing the thread. Mistral also says reasoning effort i(mistral.ai)er, deeper run. (mistral.ai) ### Why does the 256k context window matter? Because agent workflows are context hogs. A normal chat can survive on a short memory. A coding or research agent cannot. It needs room for codebase chunks, prior tool calls, user instructions, error traces, and intermediate notes. A 256k window does not magically solve reliability, but it gives the model enough working space to keep whole documents and long sessions in view instead of constantly summarizing itself into confusion. That is the practical appeal here. (mistral.ai) ### Are there real performance signals yet? Mistral’s own release materials point to 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom, with claims that the model runs strongly on agentic tasks and structured outputs. Those are promising numbers, especially because the company is tying them to a concrete product launch rather than just a benchmark table. But the catch is that, right now, most of the evidence is still coming from Mistral itself. Independent comparisons will matter. (mistral.ai) ### What changed from Medium 3? The shift is less about a clean benchmark jump and more about product posture. Medium 3, released in May 2025, was framed around enterprise efficiency and lower-cost deployment. Medium 3.5 turns that same lane into an agent story — open weights, long-horizon tasks, cloud execution, and a default spot inside Mistral’s own tools. In other words, Mistral is moving from “here is a capable model” to “here is the model running the workflow.” (mistral.ai) ### Why bundle it with remote agents? Because that is where the market is going. A coding assistant that waits for every user nudge is useful, but limited. A remote agent that keeps running in parallel while you step away is much closer to actual labor automation. Mistral says Vibe sessions can run in the cloud, in parallel, and even be started from Le Chat or “teleported” from the CLI. Medium 3.5 is the model that supposedly makes those longer runs practical enough to ship. (mistral.ai) ### So what is the bottom line? This release matters less as a raw model event than as a packaging event. Mistral is trying to prove that a single open-weight-ish, self-hostable model can sit underneath real agent products and hold up over long tasks. If that works, Medium 3.5 is not just another entry in the benchmark race. It is Mistral’s argument that the next competition is about who can turn models into dependable workers. (mistral.ai)