Mistral Unveils Medium 3.5 and Adds Agent Mode to Le Chat

Mistral AI announced on May 2 a new flagship model and a substantial upgrade to its Le Chat assistant. Medium 3.5, a 128-billion-parameter model, combines chat, reasoning, and coding into a single system, while Work mode brings agentic behaviour to the consumer-facing chatbot.

The Paris-based lab posted a SWE-Bench Verified score of 77.6 percent for Medium 3.5, placing it within range of the leading coding models from OpenAI, Anthropic, and Google. Mistral describes the model as its first to fold the three core capabilities into one architecture, rather than offering separate chat and reasoning variants.

Work mode and connectors

The more visible change for users is Work mode in Le Chat. Powered by Medium 3.5, the agent can call multiple tools in parallel and work through multi-step projects until they finish. Mistral has built in connectors to email and calendar systems so the assistant can summarise inboxes, schedule meetings, or pull together a research brief without the user steering each step.

Sensitive actions still require explicit user approval. Sending a message, deleting a calendar entry, or booking on behalf of a user pauses for confirmation. The intent, according to Mistral's product notes, is to keep the agent useful for office work without surrendering control of the inbox.

Remote agents for developers

Alongside the consumer launch, Mistral added remote agents to its developer product, Vibe. These run in the cloud and keep working through long coding sessions independently of the user's machine. The same Medium 3.5 model sits underneath.

The release lands in a crowded month. Mistral, which has long positioned itself as the European alternative to American labs, is using the 3.5 line to argue that it can stay in the same conversation as OpenAI, Anthropic, and Google.

Sources: Mistral AI announcement, TestingCatalog, The Decoder.