原文 · 未翻译
Mistral's new flagship Medium 3.5 folds chat, reasoning, and code into one model
Key Points
Mistral has released Medium 3.5, a 128-billion-parameter AI model that handles chat, reasoning, and coding tasks using a dense architecture, along with a toggleable reasoning feature for more complex queries.
The company's developer tool Vibe now includes asynchronous cloud agents that can independently handle routine tasks like bug fixes, running in isolated sandboxes with integrations for services such as GitHub and Slack.
Mistral's AI assistant Le Chat introduces a "work mode" for multi-step workflows, connecting directly to emails and calendars through built-in connectors while requiring explicit user approval before carrying out any sensitive actions.
Mistral's new flagship, Mistral Medium 3.5, merges what used to be separate models for chat, reasoning, and code into a single product. The French company is also adding asynchronous cloud agents to its coding tool Vibe and giving Le Chat a new agent mode.
Per the model card, Mistral Medium 3.5 is a dense model with 128 billion parameters and a 256,000-token context window. "Dense" means all 128 billion parameters get loaded and activated for every token generated. That makes inference expensive, but it's also simpler to run and tends to hold up better in production.
Mistral knows there are cheaper approaches. Mistral Large 3 uses a Mixture of Experts (MoE) setup with 675 billion total parameters but only activates 41 billion per token. Mistral Small 4 has 119 billion parameters and activates just 6 billion. Competitors like Deepseek and Qwen have been moving their top models toward MoE for a while, since it delivers cheaper inference at similar quality.
Against that backdrop, building the new flagship as a pure dense model is a conservative call: less optimized for inference cost, but easier to ship as one unified model for chat, reasoning, code, and agents.
Mistral says the model can be self-hosted on four GPUs. In practice, that's likely still out of reach for most users outside well-equipped data centers.
Reasoning becomes a toggle, new vision encoder built from scratch
The model follows the industry shift away from separate reasoning models, adding reasoning as a parameter on each query instead. A reasoning_effort setting switches between quick replies and a heavier mode for complex agent tasks. Mistral also retrained the vision encoder from scratch to handle variable image sizes and aspect ratios.
reasoning_effort
In Mistral's own benchmarks, Medium 3.5 scored 77.6 percent on SWE-Bench Verified and 91.4 percent on T3-Telecom. Mistral says the model replaces Medium 3.1 and the Magistral reasoning model in Le Chat, plus Devstral 2 in the Vibe CLI.