Mistral Medium 3.5

Mistral has unveiled Mistral Medium 3.5, a new flagship model that merges instruction-following, reasoning, and coding into a single dense 128B parameter architecture with a 256k context window. Released as open weights under a modified MIT license, the model is designed for self-hosting on as few as four GPUs, making it accessible for enterprise deployments. Mistral Medium 3.5 achieves strong benchmark results, scoring 77.6% on SWE-Bench Verified, outperforming models like Devstral 2 and Qwen3.5 397B A17B, and 91.4 on τ³-Telecom for agentic capabilities. The model supports configurable reasoning effort per request, allowing it to handle quick chat replies or complex, multi-step agentic tasks. Its vision encoder was trained from scratch to handle variable image sizes and aspect ratios.

Alongside the model, Mistral introduced remote coding agents in Mistral Vibe, which run asynchronously in the cloud. Developers can start coding sessions from the Vibe CLI or directly within Le Chat, offloading tasks without leaving their conversations. Multiple agents can run in parallel, and users can inspect progress via file diffs, tool calls, and status updates. Ongoing local CLI sessions can be teleported to the cloud with full session history and task state preserved. The agents integrate with GitHub, Linear, Jira, Sentry, and communication tools like Slack or Teams. Each session runs in an isolated sandbox, and upon completion, agents can open pull requests on GitHub and notify users. Mistral also introduced Work mode in Le Chat, a powerful agent for complex, multi-step tasks such as research, analysis, and cross-tool actions, powered by Mistral Medium 3.5.