Mistral takes coding agents to the cloud

Mistral Medium 3.5 is Mistral's new 128B dense model built for long-horizon coding and complex workflows. It ships as open weights under a modified MIT license and serves as the execution engine behind remote asynchronous agents in Mistral Vibe and Le Chat, letting you hand off coding tasks to cloud sandboxes instead of tying up your local laptop.

What it does

Until now, local coding agents meant frozen terminal tabs, spinning laptop fans, and constant babysitting. Mistral shifts that workload into the cloud. You can kick off tasks from the Mistral Vibe CLI or directly inside a Le Chat conversation, running multiple jobs in parallel while you move on to other work.

One of the best mechanics here is teleportation. If you start a coding session in your local CLI, you can beam the entire active session up to the cloud runtime. It brings along your full session history, task state, and tool permissions without skipping a beat.

Every cloud agent runs inside an isolated sandbox capable of broad file edits and package installations. Vibe connects straight to GitHub, Jira, Linear, Sentry, and Slack. It handles high-volume tasks like test generation, dependency updates, and CI debugging, and once finished, it automatically opens a GitHub Pull Request and pings you for review.

Le Chat also gains a new Work mode powered by the same model. It orchestrates cross-tool workflows across email, calendar, and documents, calling multiple tools in parallel while asking for explicit permission before sensitive operations like sending messages or editing files.

Specs, benchmarks, and how to run it

Mistral Medium 3.5 is a dense 128B parameter model packing a 256k context window. It blends instruction-following, multi-step reasoning, and coding into one set of weights, with per-request configurable reasoning effort and a custom vision encoder trained from scratch.

On the numbers front, it posts 77.6% on SWE-Bench Verified, beating Devstral 2 and large competitors like Qwen3.5 397B A17B. It also scored 91.4 on the tool-calling benchmark τ³-Telecom.

For self-hosting, the model can run on as few as four GPUs. The open weights are available on Hugging Face under a modified MIT license. If you prefer managed infrastructure, NVIDIA hosts it on build.nvidia.com and via NVIDIA NIM microservices.

Via API, Mistral charges $1.50 per million input tokens and $7.50 per million output tokens. Remote cloud agents and Work mode are live now for Pro, Team, and Enterprise users on Vibe and Le Chat.

Why it matters

If you are a developer bogged down by maintenance tasks, this setup is worth spinning up immediately. Offloading refactors, test suites, and dependency bumps to parallel cloud sandboxes that wrap up with clean Pull Requests removes the local terminal bottleneck entirely.

By keeping a 128B dense footprint deployable on four GPUs while putting up a 77.6% SWE-Bench score, Mistral shows that enterprise-grade autonomous coding does not require a monolithic trillion-parameter model. Teleporting CLI sessions to the cloud solves one of the biggest friction points in developer agents today.