Mistral releases Medium 3.5 and cloud coding agents in Vibe

Mistral has launched Mistral Medium 3.5 — a new flagship model with 128 billion parameters and cloud agents in the Vibe IDE. Now long coding tasks can be launched in the cloud and left running in the background while you work on other things.

Mistral Medium 3.5 — a model for long-horizon tasks

This is Mistral's first unified model that handles three directions in a single set of weights: instruction following, reasoning, and coding. Developers often use different models for different stages — for quick responses and for complex analysis. Medium 3.5 eliminates this fragmentation. A context window of 256 thousand tokens, can run self-hosted on four GPUs. Performance-wise: 77.6% on SWE-Bench Verified (better than Devstral 2 and Qwen 3.5), and 91.4 on τ³-Telecom for agent tasks. The model is built for long horizons with multiple tool calls and structured output, which then processes code. A new detail: reasoning intensity is now configurable per request. One model can answer a quick question or conduct deep analysis.

Cloud agents in Vibe

The main point: coding sessions now run asynchronously in the cloud. Many sessions can run in parallel, you stop being the bottleneck at each step. Start a task — it runs on its own, notifies you when ready. Can be launched from Vibe CLI or directly from Le Chat without interrupting the conversation. If there's already a local session, it can be teleported to the cloud — context, history, and approvals are fully transferred.

Here's what agents support:

• Launch from CLI or Le Chat
• Parallel execution of multiple tasks
• Work in an isolated sandbox (broad edits, installs)
• Integration: GitHub (code, PR), Linear, Jira, Sentry, Slack
• Final pull request with notification

During execution you see file diffs, tool calls, progress, and questions from the agent. When ready, it can open a PR and send a notification — you see the result, not every step.

Work mode in Le Chat

Complementing cloud agents is a new Work mode in Le Chat (in preview). Designed for complex multi-step tasks: research, analysis, cross-tool actions (when you need to check multiple sources, process results, write conclusions). Runs on Medium 3.5. The agent can invoke tools in parallel while the task runs in parallel, and outputs structured results that code can further process.

"Coding agents lived on your laptop.

Now we're moving them to the cloud, where they work independently, in parallel, and notify you when ready," Mistral says in its announcement.

What this means

Cloud agents are moving from experimental stage to production. For developers, this means less routine on parallel tasks — instead of sitting and waiting for each step, you send it to the cloud. For companies, it's a way to scale development without linearly increasing the team. And the open-weights 128B model that runs on 4 GPUs is available to anyone who wants to self-host.