Mistral moves coding agents to the cloud with Medium 3.5

Mistral AI launched Medium 3.5, remote coding agents in Vibe and Work mode in Le Chat on April 29.

The facts first: Medium 3.5 is in public preview. Mistral describes it as a 128B dense model with a 256k context window, built to combine instruction following, reasoning and coding in one model. It is released as open weights under a modified MIT license, can according to Mistral be self-hosted on as few as four GPUs, and becomes the default model in Le Chat and Vibe CLI. API pricing is listed at $1.5 per million input tokens and $7.5 per million output tokens.

For leaders, the more important move is not the model card. Mistral is moving coding agents from the local terminal into remote Vibe sessions. A developer can start work from the CLI or Le Chat, let it continue in the cloud, inspect file diffs, tool calls and progress, and receive a pull request back in GitHub. Mistral says GitHub, Linear, Jira, Sentry, Slack and Teams are part of the workflow.

That raises a practical CIO question: should agentic coding run on the developer laptop, in the vendor cloud, in the company's cloud, or in a controlled internal environment? The answer affects source-code exposure, logging, access to issue trackers, pull-request approval and who controls build and test environments.

Work mode in Le Chat extends the same pattern to knowledge work. Mistral says the agent can read and write, use several tools in parallel, and pull context from email, calendar, documents and internal sources. It also says sensitive actions such as sending a message, writing a document or modifying data require explicit approval based on the user's permissions.

Assessment: this is a European alternative to the agent stacks from OpenAI, Microsoft, Google and AWS, not just another chatbot release. Open weights and a relatively low self-hosting threshold create architecture options, but remote agents and default-enabled connectors raise the governance bar.

The concrete advice for leadership is to avoid treating this as an unmanaged developer experiment. Start with a control matrix: where the agent may read, where it may write, which repositories it may open pull requests against, who approves sensitive actions, and what logs the security team receives. If Mistral is evaluated as a supplier, procurement and security should review the data-processing agreement, open-weight license terms, model-hosting model and exit strategy before developer teams make it a default tool.