Mistral Medium 3.5 powers remote Vibe agents

Mistral Medium 3.5 has arrived in public preview. It's a unified model with a 128B dense architecture and a 256k context window, combining instruction-following, reasoning, and coding into a single set of weights. Reasoning effort can be adjusted per request, and the vision encoder was trained from scratch to handle variable image sizes and aspect ratios.

Performance is reported at SWE-Bench Verified 77.6% and τ³-Telecom 91.4. Self-hosting is possible with as few as 4 GPUs, and the model is released with open weights under a modified MIT license.

This model has become the default model for Mistral Vibe and Le Chat, replacing Devstral 2 in the Vibe CLI. Remote coding sessions run in parallel in the cloud, supporting the following flows:

• Start directly from the CLI or Le Chat
• Move a local CLI session to the cloud
• Integrations with GitHub, Linear, Jira, Sentry, Slack, Teams
• PR creation and notification handling when a task finishes

Work mode (Preview) orchestrates Workflows in Mistral Studio to handle multi-step tasks. It reviews email, messages, calendar, internal documents, and the web together, running tool calls in parallel to push through research, analysis, triage, and drafting. Connectors are enabled by default, and sensitive actions go through explicit approval based on permissions.

Pricing is $1.5 per million input tokens and $7.5 per million output tokens via API. The feature is available on Pro, Team, and Enterprise plans, and the model can also be used on build.nvidia.com and NVIDIA NIM.