Imagine coding agents no longer confined to your laptop; they are now seamlessly integrated into the cloud. These agents can operate autonomously, handling tasks in parallel and alerting you upon completion. You can easily initiate these tasks from the Mistral Vibe CLI or directly within Le Chat, allowing you to delegate coding responsibilities without disrupting your conversation.
The backbone of this innovation is Mistral Medium 3.5, now in public preview. This model serves as the new default in Mistral Vibe and Le Chat, purpose-built for extended coding and productivity tasks. Additionally, the newly introduced Work mode in Le Chat (Preview) enhances this capability, empowering agents to manage complex, multi-step tasks such as research, analysis, and cross-tool actions.
Highlights.
- Mistral Medium 3.5 is a groundbreaking flagship model that combines instruction-following, reasoning, and coding into a compact 128B dense architecture. It has been released as open weights under a modified MIT license.
- This model delivers impressive real-world performance and can operate self-hosted on just four GPUs.
- Mistral Vibe introduces remote agents for asynchronous coding, allowing you to initiate sessions from the CLI or Le Chat, and transfer ongoing local sessions to the cloud seamlessly.
- You can kickstart coding tasks in Le Chat, with sessions running on the same remote infrastructure, ensuring continuity even when you need to step away.
- The Work mode in Le Chat is driven by a new agent powered by Mistral Medium 3.5, collaborating on multi-step tasks and executing multiple tools simultaneously until objectives are met.
Mistral Medium 3.5.
Mistral Medium 3.5 marks the introduction of our first flagship merged model, currently in public preview. This dense 128B model features a 256k context window, capable of instruction-following, reasoning, and coding within a single weight set. It excels in real-world scenarios, with the option for self-hosting on a minimal four-GPU setup. The model can now configure reasoning efforts per request, enabling it to handle simple queries or complex tasks equally well. Moreover, the vision encoder was developed from the ground up to accommodate various image sizes and aspect ratios.
In performance assessments, Mistral Medium 3.5 achieved a score of 77.6% on SWE-Bench Verified, surpassing Devstral 2 and the Qwen3.5 397B A17B models. It also exhibits robust agentic capabilities, scoring an impressive 91.4 on τ³-Telecom.
This model is particularly well-suited for long-duration tasks, reliably engaging multiple tools while generating structured outputs suitable for further processing. It has paved the way for the practical deployment of asynchronous cloud agents in Vibe.
As the default model in Le Chat, Mistral Medium 3.5 also replaces Devstral 2 within our coding agent, Vibe CLI.
Vibe Remote Agents.
Starting today, coding sessions can efficiently handle lengthy tasks even while you are away. Multiple sessions can run simultaneously, removing the bottleneck that your presence often causes during each step of the process.
You can launch cloud agents via the Mistral Vibe CLI or Le Chat. As they operate, you can monitor agent activities, view file diffs, tool calls, progress states, and address any questions that arise along the way. Furthermore, you have the ability to teleport ongoing local CLI sessions to the cloud when you want them to keep running, carrying over session history, task states, and approvals.
The Vibe platform integrates seamlessly into existing systems engineering environments, ensuring that human oversight is applied where necessary. It features compatibility with GitHub for code and pull requests, Linear and Jira for issue tracking, Sentry for incident management, along with collaboration tools like Slack and Teams for updates.
Each coding session is executed within an isolated sandbox, accommodating extensive edits and installations. Once a task is completed, the agent can submit a pull request on GitHub, notifying you to review the final outcome without delving into every keystroke taken to achieve it.
This setup is ideal for high-volume, well-defined tasks that consume a developer’s time without infringing on their judgment; it covers module refactoring, test generation, dependency upgrades, CI investigations, and bug fixes.
We utilize Workflows organized within Mistral Studio to integrate Mistral Vibe into Le Chat. Originally developed for our internal coding environment and our enterprise customers, this functionality is now accessible to everyone, enabling web-based task initiation. Developers can now run multiple coding sessions simultaneously without being confined to a local terminal.
You can initiate coding tasks directly in Le Chat, allowing tasks outlined in chat to execute on the same remote runtime as the CLI and web, resulting in a finished branch or draft PR upon completion.
New Work Mode in Le Chat (Preview).
Work mode introduces a robust new agentic function for complex tasks in Le Chat, powered by a new harness and Mistral Medium 3.5. This agent acts as the execution backend for the assistant itself, allowing Le Chat to read, write, and utilize several tools concurrently, managing multi-step projects until completion.
Currently, Work mode enables you to:
- Execute cross-tool workflows, aligning various communications such as email, messages, and calendar into one streamlined process; gather meeting readiness materials, including attendee context, recent news, and discussion points from multiple sources.
- Engage in thorough research and synthesis across web sources, internal documents, and connected tools, producing structured briefs or reports that you can modify before sharing or exporting.
- Manage your inbox and draft responses; create tasks in Jira based on team and customer discussions; and summarize information for your team on Slack.
Sessions in Work mode last longer than conventional chat replies, allowing the agent to persist across multiple interactions, engage in trial-and-error, and drive tasks to completion. In this mode, connectors are enabled by default, enabling the agent to access documents, mailboxes, calendars, and other systems for the rich context necessary to make informed decisions.
All actions the agent performs are transparent; you can see every tool call and the associated reasoning. Le Chat will seek your explicit approval—according to your permissions—before undertaking sensitive actions, such as sending a message, drafting a document, or modifying any data.
Get Started.
Mistral Medium 3.5 is now available in Mistral Vibe and Le Chat, enabling remote coding agents and Work mode features in Le Chat across Pro, Team, and Enterprise plans.
The API pricing is set at $1.5 per million input tokens and $7.5 per million output tokens. Open weights can be found on Hugging Face under a modified MIT license.
Additionally, it is accessible for prototyping on NVIDIA GPU-accelerated endpoints at build.nvidia.com and as a scalable containerized inference microservice at NVIDIA NIM.
Join Us in Shaping the Future of Agentic Systems.
We are seeking talented individuals across research, engineering, and product teams to further advance agentic systems. Explore our open roles.
