Remote agents in Vibe. Powered by Mistral Medium 3.5.

Mistral AI introduced Mistral Medium 3.5 in public preview, describing it as its first flagship merged model. It is a dense 128B model with a 256k context window that handles instruction-following, reasoning, and coding in a single set of weights. It is released as open weights under a modified MIT license and can be self-hosted on as few as four GPUs, according to the company.

The model becomes the default in Mistral Vibe and Le Chat, replacing Devstral 2 in the Vibe CLI coding agent. Mistral reported that Medium 3.5 scores 77.6% on SWE-Bench Verified, ahead of Devstral 2 and models such as Qwen3.5 397B A17B, and 91.4 on τ³-Telecom. Reasoning effort is configurable per request, and the company said it trained the vision encoder from scratch to handle variable image sizes and aspect ratios.

Alongside the model, Mistral moved coding agents to the cloud through remote agents in Vibe. Sessions can be started from the Vibe CLI or Le Chat, run in parallel and asynchronously in isolated sandboxes, and notify the user when complete. Local CLI sessions can be teleported to the cloud with session history, task state, and approvals carrying across. The agents integrate with GitHub, Linear, Jira, Sentry, Slack, and Teams, and can open pull requests when work is done. Mistral said the feature suits module refactors, test generation, dependency upgrades, CI investigations, and bug fixes.

Le Chat also gains Work mode in preview, an agentic mode powered by a new harness and Medium 3.5. It runs multi-step tasks such as cross-tool workflows, research and synthesis, inbox triage, and creating Jira issues, calling tools in parallel. Connectors are on by default, and Mistral said Le Chat asks for explicit approval, based on user permissions, before sensitive actions such as sending a message, writing a document, or modifying data. Every tool call and thinking rationale is visible, it said.

Medium 3.5 is available today in Mistral Vibe and Le Chat on Pro, Team, and Enterprise plans. Through the API, it is priced at $1.5 per million input tokens and $7.5 per million output tokens. Open weights are on Hugging Face, and it is also available on NVIDIA GPU-accelerated endpoints on build.nvidia.com and as NVIDIA NIM.