Mistral Medium 3.5: Cloud Agents and a New Work Mode for Complex Tasks

Mistral AI has launched Mistral Medium 3.5, a new 128B model that combines instruction following, reasoning and coding. The main change: coding agents move to the cloud and now run asynchronously and in parallel, notifying the developer when a task is done.

Mistral Medium 3.5: a model for long-running tasks

Mistral Medium 3.5 is the first merged model in the Mistral lineup, with all capabilities in a single set of weights instead of separate models for separate tasks. At 128B parameters with a 256k context window, it can run on four GPUs, which makes self-hosting practical and affordable.

The performance is impressive. On SWE-Bench Verified the model scores 77.6%, ahead of Devstral 2. For agentic tasks, where the model has to call tools and solve multi-step problems, it scores 91.4 on τ³-Telecom. That means Medium 3.5 works reliably in long sessions with many tools and produces structured output that other programs can process.

The reasoning effort parameter is now configurable per request. The same model can answer a quick chat query in milliseconds or spend more time on a complex agentic loop with multi-step reasoning. This is flexibility that did not exist before.

Cloud agents in Vibe: asynchronous and parallel

Previously, coding agents ran locally on the developer's laptop. That meant the developer had to sit at the screen, watch every step and approve actions. Now agents move to the cloud and run asynchronously and in parallel, regardless of where the developer is.

There are two ways to launch a cloud agent. The first is from Mistral Vibe CLI: you type a command, and the session spins up in the cloud and works on its own. The second is directly from Le Chat: describe the task in the chat, and the agent picks up the context and starts working.

There is also a third option: if you started work locally in the CLI and decided to step away, you can teleport the session to the cloud. The full conversation history, task state and approvals already given carry over, so work continues smoothly.

Each session runs in an isolated sandbox and has access to the tools the specific job needs: GitHub (code and pull requests), Linear and Jira (task management), Sentry (incident monitoring), Slack and Teams (notifications and communication).

When the work is done, the agent automatically opens a pull request. You simply review the result instead of watching every step.

Work mode in Le Chat for complex tasks

The new Work mode in Le Chat launches a dedicated agent, powered by Medium 3.5, that breaks down complex multi-step tasks. These include research, data analysis and cross-tool actions across several integrations at once. The agent calls tools in parallel and keeps working until the task is complete.

The system targets high-volume, well-defined work that consumes developer time but does not require their judgment at every step:

• Refactoring large modules
• Generating tests for existing code
• Updating and fixing dependencies
• Analyzing CI problems and logs
• Fixing known bugs

All of this is routine work that demands accuracy and attention, but not creativity.

What this means

Cloud coding agents are moving out of labs and experiments into production. For engineering teams, this means they can finally delegate high-volume, well-defined tasks (refactors, tests, updates) and get a finished pull request to review. Developers no longer spend hours on mechanical work and can focus on what requires their judgment and creativity.