Mistral launches Medium 3.5

Mistral just shipped a new flagship model, but the bigger story is what kind of flagship it is. Medium 3.5 is not another closed API-only release aimed at proving a benchmark point. It is a 128B dense multimodal model with open weights, a 256k context window, and a design that tries to collapse chat, reasoning, and coding into one system. Mistral launched it on April 29, 2026, at the same time it rolled the model into Vibe remote agents and Le Chat’s new Work mode. (mistral.ai) ### What actually launched? Three things moved together. Medium 3.5 became Mistral’s new frontier model, Vibe got remote coding agents that run in the cloud instead of only on your laptop, and Le Chat got a Work mode for longer, more complex tasks. That packaging matters — Mistral is not presenting the model as a lab artifact. It is presenting it as the engine behind products people can actually use. (mistral.ai) ### What is Medium 3.5 supposed to be? Basically, it is Mistral’s “one model instead of many” play. The company describes Medium 3.5 as its first flagship merged model — a single set of weights for instruction following, reasoning, and coding, with image input and text output. It also exposes a configurable reasoning mode, so developers can trade speed for more test-time compute when a task needs it. (docs.mistral.ai) ### Why does “merged model” matter? Because Mistral had been splitting these jobs across separate families. Medium 3.5 replaces Medium 3.1 and Magistral inside Le Chat, and it replaces Devstral 2 inside Vibe. That means the company is betting users would rather have one stronger general-purpose model than bounce between a chat model, a reasoning model, and a coding model depending on the task. (huggingface.co) ### What are the headline specs? The concrete numbers are straightforward. Medium 3.5 is dense, not mixture-of-experts, at 128B parameters. It supports a 256k context window. It is multimodal, taking text and image input. Mistral lists pricing at $1.5 per million input tokens and $7.5 per million output tokens in its docs, whi(huggingface.co)mos. (docs.mistral.ai) ### How open is it, really? More open than most frontier-ish releases, but not a public-domain free-for-all. Mistral published the weights on Hugging Face and says the model is released under a Modified MIT license. That is a meaningful distinction from the big hyperscaler pattern, where the model is available only throu(docs.mistral.ai)in, this is the part that changes the conversation. (huggingface.co) ### Are the benchmark claims impressive? The eye-catching number is 77.6% on SWE-Bench Verified, plus 91.4% on τ³-Telecom in Mistral’s model card. Those are strong agentic and coding-oriented claims, and they explain why the company immediately tied the release to Vibe’s remote coding agents. The catch is that these are vendor-publi(huggingface.co) more “does it hold up in real tool-using workflows?” (huggingface.co) ### Why are people focusing on self-hosting? Because a dense 128B model with open weights is a very specific offer. It is big enough to look serious, but Mistral is also framing it as deployable infrastructure rather than pure research theater. The Hugging Face release even spawned immediate community quantizations, which is usually a sign that operators want to run it outside the original vendor stack. (huggingface.co) ### So what changed in the market? The center of gravity moved a bit toward “open enough to build on.” Mistral is trying to prove that a flagship model can still be productized, benchmarked, and integrated into agents without being locked behind a single closed endpoint. If that works, Medium 3.5 will matter less as one more model l(huggingface.co) agents on top. (mistral.ai) ### Bottom line? This is Mistral making its clearest pitch yet: you should not have to choose between a capable flagship model and the ability to run it on your own terms. (docs.mistral.ai)