Coding agents primarily reside on laptops. We are now moving them to the cloud, where they automatically run in parallel and notify us when they are complete. You can start from the Mistral Vibe CLI or directly in Le Chat, offloading your coding tasks without leaving the conversation.
Powering this is Mistral Medium 3.5, which is in public preview. This is the new default model for Mistral Vibe and Le Chat, and it’s built to last for long runs on coding and productivity tasks. A new work mode in Le Chat (preview) extends this with a powerful agent for complex multi-step tasks such as investigation, analysis, and cross-tool actions.
Highlights.
- Mistral Medium 3.5 is a new flagship model that combines the ability to follow instructions, reason, and code into one 128B high-density model. Released as open weight under a modified MIT license.
- Delivers powerful real-world performance in a size that allows it to self-host on just 4 GPUs.
- Mistral Vibe remote agent for asynchronous coding: Sessions run in the cloud and can be generated from the CLI or Le Chat, and local CLI sessions can be teleported to the cloud.
- Start your Mistral Vibe coding task with Le Chat. The session runs on the same remote runtime and continues while the user is away.
- Le Chat’s work mode runs on a new agent powered by Mistral Medium 3.5 and works through multi-step tasks, invoking tools in parallel until the job is completed.
Mistral medium 3.5.
Mistral Medium 3.5 is the first flagship integration model available in public preview. It is a dense 128B model with a 256k context window that handles instruction tracking, inference, and coding with a single set of weights. In real-world use, it has strong performance and is self-hosted with just 4 GPUs. Inference work can now be configured on a per-request basis, so the same model can respond to quick chat responses or handle complex agent execution. We trained a vision encoder from scratch to handle variable image sizes and aspect ratios.
Mistral Medium 3.5 received a score of 77.6% on SWE-Bench Verified, outperforming models such as Devstral 2 and Qwen3.5 397B A17B. It also has powerful agent features and has a score of 91.4 on τ³-Telecom.
This model is built for long-term tasks, reliably calling multiple tools and producing structured output that can be used by downstream code. This is the model that makes it practical to ship Vibe’s asynchronous cloud agents.
Mistral Medium 3.5 will be the default model for Le Chat. It also replaces the coding agent Devstral 2 for Vibe CLI.
Vibe remote agent.
Starting today, you can tackle long tasks on the go with coding sessions. Many can be run in parallel, eliminating bottlenecks in every step an agent takes.
Cloud agents can be launched from Mistral Vibe CLI or Le Chat. While it’s running, you can inspect the agent’s behavior, including file diffs, tool calls, progress, and questions that appear during progress. If you want to keep an ongoing local CLI session running, you can teleport it to the cloud, inheriting session history, task status, and approvals.
Vibe sits between the systems engineering teams already use and can collaborate with humans whenever needed. It plugs into GitHub for code and pull requests, Linear and Jira for issues, Sentry for incidents, and apps like Slack and Teams for reports.
Each coding session, including extensive editing and installation, is performed in an isolated sandbox. When the work is complete, the agent opens a pull request on GitHub and notifies you, so you can see the results instead of all the keystrokes that produced them.
This is suitable for large amounts of, well-defined work that takes time without requiring developer judgment, such as module refactoring, test generation, dependency upgrades, CI exploration, and bug fixing.
Bring a Mistral feel to Le Chat with a workflow tailored in Mistral Studio. It was initially built for internal coding environments and later for enterprise customers. This feature is now open to everyone, allowing you to launch coding tasks from the web. It also allows developers to run multiple terminals in parallel without being tied to a local terminal.
You can start a coding session directly in Le Chat, so the tasks described in the chat are executed in the same remote runtime as the CLI or web, and come back later as a completed branch or draft PR.
Le Chat’s new working mode (preview).
Work Mode is a powerful new agent mode for complex tasks in Le Chat, powered by a new harness and Mistral Medium 3.5. Because the agent becomes the execution backend for the assistant itself, Le Chat can read and write, use multiple tools at the same time, and run multi-step projects until what is requested is completed.
Here’s what you can do in Work mode today:
- Cross-tool workflow: Process email, messages, and calendars in one run. Prepare for the meeting with talking points from audience context, latest news, and information sources.
- Research and Integrate: Drill into topics across the web, internal documents, and connected tools to create structured summaries or reports that you can edit before exporting or sending.
- Prioritize your inbox and reply drafts. Create issues in Jira from team and customer discussions. Send a summary to your team in Slack.
Sessions last longer than regular chat responses, allowing agents to keep trying until they’re done. In work mode, connectors are turned on by default, rather than manually selected, so agents can access documents, mailboxes, calendars, and other systems to get the rich context they need to take the right actions.
All actions performed by the agent are displayed. The invocation of each tool and the reasoning behind it are displayed. Le Chat asks for explicit approval based on your permissions before proceeding with sensitive tasks such as sending messages, creating documents, or modifying data.
Let’s get started.
Mistral Medium 3.5 is currently available on Mistral Vibe and Le Chat, and powers remote coding agents and working modes in Le Chat on Pro, Team, and Enterprise plans.
Through the API, pricing is set at $1.5 per million input tokens and $7.5 per million output tokens. Open Weight is published on Hugging Face under the modified MIT License.
It is also available for prototyping as NVIDIA NIM, a scalable, containerized inference microservice hosted on the NVIDIA GPU-accelerated endpoint at build.nvidia.com.
Build the future of agent systems with us.
We are hiring in research, engineering, and product to further advance our agent systems. Check out our open positions.
