Mistral Medium 3.5

Mistral Medium 3.5 has been released. A single 128B dense model handles instruction-following, reasoning, and coding, and it supports a 256k context window and per-request reasoning effort.

Performance is presented as SWE-Bench Verified 77.6% and τ³-Telecom 91.4. The vision encoder was trained from scratch to handle various image sizes and aspect ratios, and it is designed for structured output and multiple tool calls.

Product integrations were announced together.

• In Vibe, coding sessions run asynchronously in the cloud, and multiple sessions can run in parallel.
• Local CLI sessions can teleport to the cloud while preserving history, task state, and approval.
• Le Chat Work mode operates as an agentic mode that uses connected tools, documents, mailboxes, and calendars, and it requests explicit approval for sensitive tasks.

Deployment and pricing are as follows.

• It becomes the default model for Le Chat, and it replaces Devstral 2 in the Vibe CLI.
• API pricing is $1.5 per 1 million tokens for input and $7.5 per 1 million tokens for output.
• Open weights are provided on Hugging Face under a modified MIT license.
• It is also hosted on NVIDIA Build and NIM.