Mistral Medium 3.5 powers remote coding agents

Mistral launched Medium 3.5 and moved its coding agents into the cloud.

Mistral AI just pushed its agent story into a more practical shape: coding work can now run remotely, in parallel, and keep going after you close the tab. The company says Mistral Medium 3.5 is the default model for Le Chat and Vibe, and it comes with a 128B dense architecture, a 256k context window, and open weights.

What Mistral actually changed

The headline is not only a new model. Mistral is pairing the model with a different working style for developers: agents that run in the cloud, continue while you step away, and report back when they finish. That matters because a lot of coding-agent work is repetitive, interrupt-heavy, and slow when it stays pinned to a local terminal.

With Vibe remote agents, a session can start in the CLI or in Le Chat, then move to the cloud runtime without losing task state, approvals, or history. Mistral says you can even “teleport” an ongoing local CLI session into the cloud. That is a small phrase with a big implication: the company wants the same workflow across laptop, browser, and background execution.

The practical use cases are the ones teams already know well: module refactors, test generation, dependency upgrades, CI investigations, and bug fixes. Those are the jobs that eat hours because they involve many small edits and checks rather than one clever insight.

• Sessions can run in parallel instead of one at a time.
• Agents can open pull requests on GitHub when they finish.
• File diffs, tool calls, and progress states stay visible during the run.
• Local sessions can move to the cloud with state intact.

Why Medium 3.5 matters more than the branding

Hugging Face often gets crowded with model launches, but Medium 3.5 has a few details that make it more interesting than a standard flagship release. It is a single dense model that combines instruction-following, reasoning, and coding rather than splitting those abilities across separate weights. Mistral also says reasoning effort is configurable per request, which should help the same model act like a fast chat assistant in one moment and a slower agent in the next.

The benchmark numbers are solid enough to matter in real engineering conversations. Mistral says Medium 3.5 scores 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom. It also claims self-hosting on as few as four GPUs, which is important for teams that care about deployment control and cost ceilings.

“Today we’re moving them to the cloud, where they run on their own, in parallel, and notify you when they’re done.” — Mistral AI

That quote captures the product shift better than any marketing slide. Mistral is betting that developers do not want another chat box that writes code. They want a worker that can stay busy while they do something else, then hand back a branch or a draft PR instead of a stream of partial thoughts.

Work mode turns Le Chat into an execution layer

The second half of the release is Le Chat Work mode, which Mistral says is powered by a new harness plus Medium 3.5. In practice, this makes the assistant more like an execution layer than a simple chatbot. It can read and write across connected tools, call several tools at once, and keep working through multi-step tasks until the job is done.

That matters for office work as much as coding. Mistral highlights workflows like catching up across email, messages, and calendar in one run; preparing for meetings with attendee context and recent news; triaging inboxes; drafting replies; creating Jira issues; and sending summaries to Slack or Teams. The company also says connectors are on by default in Work mode, so the agent can pull in the context it needs without a manual setup step every time.

There is a real tradeoff here. More autonomy means more speed, but it also increases the need for visibility and approvals. Mistral says every action is visible, with tool calls and reasoning shown to the user, and sensitive actions still require explicit approval based on permissions. That is the right constraint if this is going to be used inside real teams.

• GitHub integration covers code and pull requests.
• Linear and Jira connect issue tracking to agent work.
• Sentry plugs incident data into the workflow.
• Slack and Microsoft Teams handle reporting and notifications.

The pricing and access details tell you who this is for

Mistral is making Medium 3.5 available today in Vibe and Le Chat on Pro, Team, and Enterprise plans. Through the API, it costs $1.5 per million input tokens and $7.5 per million output tokens. Open weights are on Hugging Face under a modified MIT license, and the model is also available through NVIDIA Build and