Mistral ships Medium 3.5

Mistral just shipped two things at once — a new flagship-ish model and a new way to use it. Medium 3.5 is the model. Remote agents in Vibe are the product wrapper. The pitch is simple: stop making coding agents live only on your laptop, move them into the cloud, and let them keep working while you walk away. Mistral launched both on April 29, with Medium 3.5 in public preview and Vibe remote agents tied directly to it. (mistral.ai) ### What is Medium 3.5, exactly? It’s a 128B dense multimodal model with a 256k context window, and Mistral is positioning it as a single merged system for instruction-following, reasoning, and coding rather than separate specialized checkpoints. The company also says it released the weights under a modified MIT license and that the model can be self-hosted(mistral.ai)es that want more control. (mistral.ai) ### Why pair it with coding agents? Because Mistral is trying to make “agentic coding” feel less like a demo and more like a workflow. The new remote agents in Vibe run in the cloud, can work in parallel, and keep going after you close the terminal. You can start a task from the Vibe CLI or from Le Chat, and Mistral even frames one feature as “teleporting” (mistral.ai)because it’s the engine that makes longer, multi-step coding runs practical. (mistral.ai) ### What’s the headline performance claim? The number Mistral wants you to notice is 77.6% on SWE-Bench Verified. It also highlights strong tool use and structured output, plus a configurable `reasoning_effort` setting so developers can trade speed for deeper reasoning on a per-request basis. That last bit is important — Mistral is pushing one model that ca(mistral.ai) the job. (mistral.ai) ### So where’s the catch? Price. Medium 3.5 is listed at $1.50 per million input tokens and $7.50 per million output tokens. That’s a sharp jump from Mistral Medium 3.1, which is listed at $0.40 input and $2 output. So even if the new model is better suited to coding agents, the economics got materially worse. For teams doing lots of long agent runs, output pricing is the part that bites. (docs.mistral.ai) ### Why would anyone still want it? Because the real differentiator may not be raw API price. Mistral has been leaning for a while into the idea that “medium” models can be good enough for serious enterprise work while being easier to deploy. Medium 3, back in May 2025, was pitched as frontier-ish performance at much lower cost a(docs.mistral.ai)lities and agent workflows. In other words — Mistral wants to own the “practical flagship” lane. (mistral.ai) ### Why does the open-weights angle matter here? Because cloud agents are useful, but plenty of companies still want the option to run sensitive workloads themselves. Open weights plus a four-GPU deployment target gives Mistral a cleaner enterprise story than vendors that only offer hosted access. The catch is that open doesn’t automatically mean cheap — hardware, orchestration, and long-running agent jobs still cost money. (mistral.ai) ### Is this really a model launch story? Partly. But mostly it’s a workflow story. Mistral is saying the next fight is not just whose benchmark bar is tallest. It’s whose model can sit inside a coding tool, call tools reliably, run for a long time, and fit into enterprise infrastructure without drama. Medium 3.5 matters because it’s the model Mistral thinks can carry that whole stack. (mistral.ai) ### Bottom line? Mistral didn’t just release another model. It bundled a model, an agent product, and a deployment argument into one message. The upside is clear — stronger coding workflows, async execution, and self-hosting flexibility. But the company also made the price question impossible to ignore. (mistral.ai)