Mistral AI has announced a major infrastructure upgrade to its coding agent platform, Vibe, introducing remote agents that can run tasks in the cloud. Alongside this, the company has released Mistral Medium 3.5, a 128-billion parameter dense model now powering both Vibe and its consumer assistant, Le Chat.
Remote Agents: Coding Without the Wait
Vibe is Mistral's command-line coding agent that automates software tasks such as writing code, refactoring modules, generating tests, and investigating CI failures. Previously, Vibe sessions were tethered to a user's local machine. With the new remote agent capability, developers can launch tasks and let them execute in the cloud, freeing their local environment.
Key features include:
- Parallel execution: Multiple sessions can run simultaneously.
- Persistent state: Local sessions can be "teleported" to the cloud, preserving history and approvals.
- Isolated sandboxes: Each session runs in a secure, isolated environment.
- GitHub integration: Agents can automatically open pull requests upon completion.
Mistral Medium 3.5: A Performance Leap
Mistral Medium 3.5 is a dense 128B model that achieves a 77.6% score on the SWE-Bench Verified benchmark, demonstrating strong software engineering capabilities. The model is available in public preview on Hugging Face and serves as the default in both Vibe and Le Chat.
Vibe + Le Chat: A Unified Agentic Framework
Remote agents in Le Chat are powered by Mistral Studio's orchestration layer, which was initially developed for internal use and enterprise clients. This architecture allows users to start coding tasks from Le Chat and monitor progress via file diffs, tool calls, and status updates.
Vibe integrates with GitHub, Linear, Jira, Sentry, and communication tools like Slack and Teams, enabling end-to-end workflow automation.
Availability
Remote agents in Vibe and Mistral Medium 3.5 are available now. Developers can access the model via Hugging Face and start using remote agents through the Vibe CLI or Le Chat.