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

Mistral AI launched Mistral Medium 3.5—a new 128B model that combines instruction-following, reasoning, and coding capabilities. The main innovation: coding agents move to the cloud and now work asynchronously and in parallel, notifying the developer when they complete a task.

Mistral Medium 3.5: A Model for Long Tasks

Mistral Medium 3.5 is the first merged model in the Mistral lineup, where all capabilities reside in a single set of weights instead of using different models for different tasks. With 128B parameters and a 256k context window, it can run on four GPUs, making self-hosting practical and cost-effective.

Performance is impressive. On SWE-Bench Verified benchmarks, the model scores 77.6%—higher than Devstral 2. For agentic tasks, where the model must invoke tools and solve multi-step problems, it scores 91.4 on τ³-Telecom. This means Medium 3.5 reliably works in long sessions with multiple 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 cycle with multi-step reasoning. This is flexibility that wasn't available before.

Cloud Agents in Vibe: Asynchronous and Parallel

Previously, coding agents worked locally on the developer's laptop. This meant the developer had to sit at the screen, monitor every step, and approve actions. Now agents move to the cloud and work asynchronously and in parallel, regardless of where the developer is located.

There are two ways to launch a cloud agent. First way: from Mistral Vibe CLI—you type a command, and the session spins up in the cloud and works independently. Second way: directly from Le Chat—write a task in the chat, and the agent understands the context and starts working. There's also a third option: if you started work locally in CLI and decide to leave, you can teleport the session to the cloud. The entire conversation history, task state, and approvals made transfer with it—continuation of work will be seamless.

Each session runs in an isolated sandbox and has access to the tools needed for the specific job: GitHub (code and pull-requests), Linear and Jira (task management), Sentry (incident monitoring), Slack and Teams (notifications and communication). When the work is complete, the agent automatically opens a pull-request. You simply review the result instead of monitoring every step.

Work Mode in Le Chat for Complex Tasks

The new Work mode in Le Chat launches a specialized agent (powered by Medium 3.5) that breaks down complex multi-step tasks. This includes research, data analysis, and cross-tool actions with multiple integrations simultaneously. The agent invokes tools in parallel and works until the task is fully completed. The system targets high-volume, well-defined work that takes developer time but doesn't require their judgment at each step:

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

All of this is routine work that requires precision and attention, but not creativity.

What This Means

Cloud coding agents are moving from labs and experiments into production. For engineering teams, this means you can finally delegate large, well-defined tasks (refactors, tests, updates) and get a ready-made pull-request for review. The developer doesn't spend hours on mechanical work but can focus on what requires their judgment and creativity.