Mistral Medium 3.5 launches with dense 128B model and remote coding agents in Vibe

Mistral has introduced Medium 3.5, its new flagship dense AI model that consolidates chat, reasoning, coding, and agentic functions within a single system. Unlike its previous approach of deploying separate specialized models, Mistral now presents unified reasoning capabilities, replacing Medium 3.1, Magistral, and Devstral 2 in core products like Le Chat and the Vibe CLI.

Medium 3.5 employs a dense architecture with 128 billion parameters and features a 256,000-token context window. This design allows it to handle extended documents, large codebases, and long multi-step workflows. While the industry is increasingly moving to Mixture of Experts models for inference efficiency, Mistral’s decision to use a dense model supports straightforward deployments but demands more hardware, namely four GPUs for self-hosting—an option mostly suited to teams and data centers. The model adds a reasoning_effort parameter, permitting users to toggle between faster or more rigorous outputs per query. For vision tasks, Mistral retrained the encoder from scratch, enabling support for images with varying sizes and aspect ratios.

Mistral also updates its tools: Vibe now features asynchronous cloud agents that can run independently, migrate local sessions to the cloud, and open pull requests, while Work Mode in Le Chat handles multi-step tasks through connectors for systems like mailboxes, calendars, and documents, while still requiring explicit confirmation for sensitive actions. Medium 3.5 weighs in with new licensing—a Modified MIT license—and is available via API and download.