Remote agents in Vibe. Powered by Mistral Medium 3.5.
Mistral AI
Mistral launched cloud-based coding agents plus Medium 3.5, a new open-weight 128B model, in Vibe and Le Chat. Agents now run async in parallel and even open pull requests without you babysitting each step.
Based on reporting by Mistral AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mistral just gave its coding agents somewhere else to live besides your laptop. Starting today, Vibe agents can run in the cloud, kicked off from the CLI or straight from a Le Chat conversation, and they'll keep working while you go do literally anything else. You get pinged when they're done. That's the pitch, anyway, and it's built on a new model called Mistral Medium 3.5, a 128B dense model that Mistral is releasing as open weights under a modified MIT license.
What's notable here isn't just the cloud move, it's the model doing the moving. Medium 3.5 merges instruction-following, reasoning, and coding into one set of weights, with a 256k context window and configurable reasoning effort depending on whether you want a quick answer or a long agentic grind. Mistral says it hits 77.6% on SWE-Bench Verified, putting it ahead of Devstral 2 and Qwen3.5's 397B A17B variant, and it scores 91.4 on τ³-Telecom for agentic tasks. It's also small enough to self-host on four GPUs, which for a model claiming flagship-tier coding performance is a genuinely useful number, not a marketing footnote.
The actual product change is the async agent workflow. Sessions run isolated, in sandboxes, with file diffs and tool calls visible as they go, and a local CLI session can be teleported to the cloud mid-task without losing history. Vibe plugs into GitHub, Linear, Jira, Sentry, Slack, Teams — the usual enterprise glue — and when a job's finished, the agent opens a pull request and steps back, letting a human review the output rather than supervise every line written. Mistral is positioning this for the grindy, well-defined stuff: refactors, dependency bumps, test generation, CI debugging, the work that eats a developer's day without needing their judgment on every move.
Alongside that, Le Chat gets a new Work mode, also running on Medium 3.5, aimed at multi-step tasks beyond code: triaging your inbox, drafting Jira issues from a Slack thread, pulling together a meeting brief from your calendar and inbox and the open web. Connectors are on by default here rather than picked manually, so the agent can reach into your documents and mailboxes without you flipping switches first — though Mistral says sensitive actions like sending a message or editing a file still require explicit approval. Pricing lands at $1.5 per million input tokens and $7.5 per million output, available now on Pro, Team, and Enterprise, with the open weights also up on Hugging Face and NVIDIA's NIM for anyone who wants to run it themselves.
My take — AI-written commentary, not fact-checked reporting
An open-weight model that's actually competitive on SWE-Bench and cheap enough to self-host on four GPUs is the real story here, not the cloud-agent UX, which every lab is racing to ship anyway. Mistral keeps proving that European open-weight releases don't have to trail the closed frontier by much, and that matters more for who controls agentic infrastructure long-term than whether your PR gets opened five minutes faster.
Read more about this at: Mistral AI