Remote agents in Vibe, powered by Mistral Medium 3.5

Mistral AI is releasing both a model and product capabilities at once this time: Mistral Medium 3.5 enters public preview and becomes the new default model for Mistral Vibe and Le Chat. The company describes it as a 128B dense flagship merged model with a 256k context window, placing instruction-following, reasoning, and coding within the same set of weights, and supporting per-request adjustment of reasoning effort. Mistral stresses that the model is designed for long-running tasks, stable invocation of multiple tools, and producing structured output that can be consumed by downstream programs, which also makes Vibe's asynchronous cloud agents feasible. The company notes it scores 77.6% on SWE-Bench Verified and can be self-hosted on as few as four GPUs; the open weights are released under a modified MIT license, with API pricing of 1.5 USD per million input tokens and 7.5 USD per million output tokens. On the product side, Mistral Vibe adds remote agents, allowing coding sessions to be launched from the Vibe CLI or Le Chat and continue executing in an isolated cloud sandbox, supporting parallel execution of multiple tasks, viewing file diffs, tool calls, progress, and issues, and also teleporting a local CLI session to the cloud while preserving history, state, and approvals. Once a task is complete, the agent can open a GitHub pull request and notify the user, making it well suited for clear but time-consuming engineering work such as module refactoring, test generation, dependency upgrades, CI investigations, and bug fixes. Le Chat also adds a Work mode Preview, powered by Mistral Medium 3.5, capable of multi-step tasks across email, messages, calendar, documents, the web, and internal tools — performing research, organization, triage, creating Jira issues, writing summaries, and sending them to Slack and the like; connectors are available by default, but sensitive actions involving sending messages, writing documents, or modifying data will require explicit approval based on permissions. Overall, this is a significant product update in which Mistral strings together model capabilities, a coding agent, and a general office agent into a single long-task workflow.