In today’s tech landscape, coding agents have primarily been anchored to our laptops. This paradigm is shifting as we embrace a more dynamic and efficient way to code by moving these agents to the cloud. Now, coding tasks can be initiated remotely, allowing agents to run autonomously in parallel and notify users upon completion. This seamless integration can be realized through the Mistral Vibe CLI or directly within Le Chat, enabling you to offload coding tasks without interrupting your conversation flow.
At the heart of this innovation is Mistral Medium 3.5, which is currently in public preview. This advanced model serves as our new default within both Mistral Vibe and Le Chat. It is purposefully designed to execute long-duration coding and productivity tasks with utmost ease. Furthermore, the newly introduced Work mode in Le Chat (also in preview) enhances these capabilities by employing a sophisticated agent adept at handling complex, multi-step tasks such as research, analysis, and cross-tool operations.
Highlights of Mistral Medium 3.5
- The centerpiece of the update is Mistral Medium 3.5, our flagship model. This 128 billion parameter dense model consolidates instruction-following, reasoning, and coding into a single entity, released with open weights under a modified MIT license.
- Real-world performance is strong, with the model capable of being self-hosted on just four GPUs, making it an appealing choice for developers.
- Mistral Vibe introduces remote agents for asynchronous coding, allowing sessions to run in the cloud. Users can initiate tasks from the CLI or Le Chat, teleporting ongoing local CLI sessions to the cloud with all their history intact.
- Initiating coding tasks in Le Chat keeps everything connected, allowing the remote runtime to function while users attend to other matters.
- Work mode in Le Chat harnesses the power of Mistral Medium 3.5, enabling the completion of complex multi-step tasks and the ability to call tools in parallel until the task is finished.
What is Mistral Medium 3.5?
Mistral Medium 3.5 represents an evolution in model architecture as our first flagship merged model available in public preview. This 128B model features a remarkable 256k context window, adeptly managing instruction-following, reasoning, and coding tasks all under one roof. The model excels in real-world applications and can be self-hosted with a minimal hardware requirement. We engineered the vision encoder from scratch to accommodate various image sizes and aspect ratios, optimizing its versatility.
The performance metrics are impressive: Mistral Medium 3.5 boasts a score of 77.6% on SWE-Bench Verified, surpassing competitors like Devstral 2 and Qwen3.5 397B A17B models. In addition, it achieves an impressive score of 91.4 on τ³-Telecom, showcasing its agentic capabilities.
The robustness of the model makes it particularly suitable for long-horizon tasks, capable of reliably integrating multiple tools to generate structured outputs that can be readily consumed by downstream code. This architecture has been crucial in making asynchronous cloud agents in Vibe a practical reality.
Vibe Remote Agents
With the introduction of Vibe, coding sessions can now tackle lengthy tasks while you focus on other priorities. Multiple agents can operate in parallel, effectively removing you as the bottleneck from processing each step. These cloud agents can be launched from either Mistral Vibe CLI or Le Chat, where you’ll have visibility into the agent’s actions, including file diffs, tool calls, and progress updates. Additionally, ongoing local CLI sessions can be transferred seamlessly to the cloud, preserving historical context and task state.
Vibe integrates directly with existing systems developers use daily, allowing human intervention where necessary. It bridges the gap with platforms like GitHub for code management, Linear and Jira for tracking issues, Sentry for incident management, and collaboration tools like Slack and Teams for updates. Each coding session operates within an isolated sandbox environment, which includes the ability to perform broad edits and installs. At the completion of a task, the agent can create a pull request on GitHub, alerting you to review the outcome without having to comb through every keystroke.
This configuration suits a high volume of well-defined developer work—think module refactors, test generation, dependency upgrades, continuous integration evaluations, and bug fixes.
Utilizing Workflows orchestrated through Mistral Studio, Vibe has been developed initially for internal use, subsequently extending capabilities to enterprise customers. This opportunity is now available for all users to initiate coding tasks via the web, linking coding actions with both CLI and web environments for complete flexibility.
Coding sessions can be launched directly in Le Chat, where tasks articulated during the conversation can seamlessly transition to the cloud, returning as completed branches or draft pull requests later.
Introducing Work Mode in Le Chat (Preview)
In a major leap forward, the new Work mode in Le Chat provides an enhanced agentic capability tailored for intricate tasks. Powered by the new harness and Mistral Medium 3.5, it functions as the execution backend for the assistant, enabling it to read and write while utilizing multiple tools concurrently to handle complex tasks until completion.
Here’s a glimpse of what Work mode enables you to accomplish today:
- Cross-tool workflows that allow you to catch up on emails, messages, and calendars in a consolidated process, along with preparing for meetings by aggregating context and relevant information.
- Research and synthesis capabilities, diving deep into topics across the web and internal documents to generate comprehensive briefs or reports ready for refinement before sharing.
- Efficient inbox triage and drafting replies, creating actionable items in Jira from discussions, and summarizing communication for your team on Slack.
The agent in Work mode persists longer than traditional chat interactions, allowing it to navigate complexities and iterative problem-solving while ensuring that every action taken is transparent. Users can view tool calls and the rationale behind decisions, with explicit permissions required for sensitive actions like sending messages or modifying data.
Getting Started with Mistral Medium 3.5
Today, Mistral Medium 3.5 is available across both Mistral Vibe and Le Chat, powering remote coding agents and the new Work mode in Le Chat for Pro, Team, and Enterprise plans. Through the API, the pricing structure is set at $1.5 per million input tokens and $7.5 per million output tokens. Additionally, open weights are hosted on Hugging Face under a modified MIT license.
For those interested in prototyping, Mistral Medium 3.5 is hosted on NVIDIA GPU-accelerated endpoints, available for exploration on build.nvidia.com. It can also be accessed as a scalable containerized inference microservice via NVIDIA NIM.
Join Us at the Frontiers of Coding Agents
We are on the lookout for talented individuals across research, engineering, and product roles to push the boundaries of agentic systems. Explore our current openings and consider becoming part of this revolutionary journey.
