Mistral AI Introduces Remote Agents to Vibe and Mistral Medium 3.5 with 77.6% SWE-Bench Verified Score

Mistral AI has been quietly building one of the first open source/heavyweight AI coding agent systems, and is shipping its most significant infrastructure upgrade yet. The Mistral team announced remote agents on Vibe, its agent coding platform, alongside the public preview of Mistral Medium 3.5 – a new compact 128B model that now serves as the default model for both Vibe and Le Chat, Mistral’s consumer assistant.

What is Vibe, and Why is it Important?

If you haven’t already used it, Mistral Vibe is a CLI (command line interface) accessible coding agent that allows an AI model to run software tasks on your behalf – writing code, refactoring modules, generating tests, investigating CI failures, and more. Think of it as a little tireless developer who can work on your entire codebase.

Until now, Vibe sessions ran locally, meaning the agent was tied to your laptop and terminal. That is changing today.

Remote Agents: Agent Runs While Away

So, now coding sessions can run on long tasks while you’re away. Many can run in parallel, and you stop being a bottleneck in every action taken by the agent.

This is an important behavioral change. Instead of monitoring the coding time on your terminal, you start the job and let the cloud handle the rest. You can start cloud agents from the Mistral Vibe CLI or Le Chat. While they are running, you can check what the agent is doing, with file variations, tool calls, progress conditions, and queries that are displayed as you go.

One very useful feature for developers who are already in the middle of a session: ongoing on-premises CLI sessions can be teleported to the cloud if you want to leave them running, with session history, task status, and permissions transferred. So you don’t lose your place — just remove the job from your machine.

Each time goes by itself. Each coding session runs in a remote sandbox, including extensive programming and installation. Once the job is done, the agent can open a pull request on GitHub and notify you, so you can update the result instead of the entire key you generated.

It is also worth understanding the concept of how Vibe connects to Le Chat. Mistral uses Workflows programmed in Mistral Studio to bring the Mistral Vibe to Le Chat – originally built for its in-house coding environment, then for business customers, and now open to everyone. This means that the remote recording agent in Le Chat is not a standalone feature – it’s built on top of the Mistral orchestration layer, which is a useful framework if you’re thinking about how to build similar agent programs yourself.

On the integration side, Vibe connects to GitHub for code and pull requests, Linear and Jira for issues, Sentry for events, and apps like Slack or Teams for reporting.

Mistral Medium 3.5: The Model Behind It All

None of this would be possible without an active AI model. This new model released by Mistral Medium 3.5the Mistral team describes it as its first integrated flagship model.

It’s a dense 128B model with a 256k core window, handling instruction-following, reasoning, and coding in a single set of weights. In context, a 256k context window means that the model can process almost 200,000 words in a single pass – long enough to think through the entire large codebase.

The model is also multimodal. The Mistral team trained the vision encoder from scratch to handle a variety of image sizes and aspect ratios — a remarkable choice of architecture. Many visual language models also use pre-trained encoders such as CLIP, so building this component from scratch suggests Mistral’s prioritized flexibility in how the model handles real-world image input rather than defaulting to fixed-resolution inference.

The Mistral Medium 3.5 scores 77.6% on SWE-Bench Verified, ahead of the Devstral 2 and models like the Qwen3.5 397B A17B. SWE-Bench Verified is a standard benchmark that tests whether a model can solve real-world GitHub problems from popular open source repositories – it’s one of the most reliable proxies of practical software engineering skill. The model also scores 91.4 on τ³-Telecom and has a strong agent potential.

One very interesting design option: the reasoning effort is now adjustable per request, so the same model can respond to a quick response to a conversation or work with a complex agent’s functionality. This is important for developers who integrate the model with an API – you can dial down the computer for a simple lookup and dial up for multi-step reasoning tasks, without changing the models.

The model is designed for long-term operations, reliably calls multiple tools, and produces structured output that can be consumed by the underlying code.

Mode of Operation in Le Chat: New Agentic Layer

Beyond the code agent improvements, Mistral is also shipping Task Mode to Le Chat — a new agent mode for standard, multi-step tasks. Task mode is a powerful new agent mode for complex tasks in Le Chat, powered by a new harness and Mistral Medium 3.5. The agent becomes the back end of the assistant itself, so that Le Chat can read and write, use several tools at once, and work on multi-step projects until it completes your request.

In practice, this means things like workflows for various tools – holding on to overall emails, messages, and calendar; preparing a meeting with relevant content drawn from multiple sources; or checking the inbox and creating Jira issues in group discussions.

In Work Mode, connectors are opened automatically rather than manually selected, allowing the agent to access documents, mailboxes, calendars, and other systems with the rich context they need to take the right action. This is a significant change in usability from traditional chat assistants, where you choose tools before each session.

Transparency is a built-in feature instead of an afterthought: every action an agent performs is visible — you see the call for each tool and the reasoning behind it. Le Chat will ask for explicit authorization – based on your permissions – before proceeding with sensitive operations such as sending a message, writing a document, or modifying data.

Key Takeaways

Here are some important things to take away:

• The Mistral Medium 3.5 is now the default model for both Vibe and Le Chat — a dense 128B model with a 256k core window that scores 77.6% on SWE-Bench Verified, beating Devstral 2 and Qwen3.5 397B A17B, and available as open weights on Hugging Face.
• Vibe code agents now run in the cloud – sessions can be started from CLI or Le Chat, run concurrently in isolated sandboxes, and local sessions can be teleported to the cloud without losing session history or work state.
• Le Chat’s new Work Mode brings multi-step interactive agent functionality – powered by Mistral Medium 3.5, it can work across e-mail, calendar, documents, Jira, and Slack at the same time, with all tool calls and virtual thinking steps and clear authorization required before critical actions.
• The think effort in Mistral Medium 3.5 is configurable per API request — the same model handles lightweight chat responses and complex long-horizon runs.

Check it out Model weights in HF again Technical details. Also, feel free to follow us Twitter and don’t forget to join our 130k+ ML SubReddit and Subscribe to Our newspaper. Wait! are you on telegram? now you can join us on telegram too.

Need to work with us on developing your GitHub Repo OR Hug Face Page OR Product Release OR Webinar etc.? contact us