Mistral Vibe: Remote Agents Powered by Mistral Medium 3.5
Mistral VibeMistral Vibe introduced remote cloud-based coding agents powered by the newly released Mistral Medium 3.5 model, enabling developers to launch async coding sessions from the CLI or Le Chat that run in isolated cloud sandboxes without requiring continuous oversight. Mistral Medium 3.5 is a dense 128B model with a 256k context window that consolidates instruction-following, reasoning, and coding into a single set of weights, scoring 77.6% on SWE-Bench Verified. Local CLI sessions can be "teleported" to the cloud while preserving conversation history and state, and multiple agents can execute in parallel to eliminate the developer as a sequential bottleneck. The model ships as open weights under a modified MIT license and is now the default model in Mistral Vibe.
Featured Video
A video we selected to help illustrate this changelog
Sources & Mentions
5 external resources covering this update
Remote agents in Vibe. Powered by Mistral Medium 3.5.
Youtube
Mistral Medium 3.5
Hacker News
Mistral AI unveils Medium 3.5 model and Work Mode for Le Chat
Testing Catalog
The French-made AI model 'Mistral Medium 3.5,' which surpasses Claude Sonnet 4.5, has been released as an open model
GIGAZINE
Mistral Launches Medium 3.5 and Cloud-Based Remote Agents for Async Coding
Open Data Science
Remote Agents Come to Mistral Vibe
Mistral Vibe introduced one of its most significant capability leaps to date: remote cloud-based coding agents that execute asynchronously, in isolation, and in parallel β all powered by the new Mistral Medium 3.5 model. With this release, developers are no longer required to stay present and supervise every step of a coding session.
Launch a Task, Walk Away
Previously, Mistral Vibe operated as an interactive local agent β the developer drove the session turn by turn. Remote agents change this model entirely. A developer can now launch a coding task from either the Mistral Vibe CLI or directly from Le Chat, and the agent will work through the task independently in a cloud sandbox. When the task completes, the developer is notified and can review the output, often in the form of a pull request ready for review.
Multiple remote agents can run simultaneously, meaning teams can parallelize work across many ongoing tasks without each one blocking the others. This shift from synchronous to asynchronous AI-assisted development is a structural change in how Mistral Vibe fits into a developer's workflow.
Teleport: From Local to Cloud
A key enabling mechanism is the Teleport feature. If a developer has already started a session locally in the Mistral Vibe CLI, they can teleport it to the cloud, which preserves the full conversation history and state. The v2.9.2 patch that followed refined this further, with Teleport surfacing the latest GitHub connection status in real time while a session is being migrated.
Mistral Medium 3.5 as Default Model
Powering these remote agents β and now set as the default model in Mistral Vibe β is Mistral Medium 3.5, a 128B dense model with a 256k context window. It replaces the previous three-model setup (Medium 3.1, Magistral, and Devstral 2) by consolidating instruction-following, reasoning, and coding into a single set of weights.
Key performance metrics:
- SWE-Bench Verified: 77.6% β measuring real GitHub issue resolution via generated code patches
- ΟΒ³-Telecom agentic score: 91.4 β testing tool selection and multi-step task execution
- Self-hostable on as few as four GPUs
The model supports configurable reasoning effort per request, enabling it to handle both quick replies and complex long-horizon tasks. API pricing is $1.5 per million input tokens and $7.5 per million output tokens.
Open Weights
Mistral Medium 3.5 is released as open weights under a modified MIT license, available on Hugging Face. It can be run self-hosted at Q4 quantization with approximately 70 GB of VRAM β territory accessible to high-end workstations and approaching consumer-grade hardware. It is also available through NVIDIA endpoints and the Mistral API.
Cloud Integrations
Remote agents in Vibe connect natively to GitHub for code and pull requests, Linear and Jira for issue tracking, Sentry for incident response, and Slack or Teams for reporting. This positions Mistral Vibe as a more complete async development platform rather than a purely interactive CLI tool.