What happened
Mistral AI released Mistral Medium 3.5, a 128B dense model featuring a 256k context window, designed for instruction-following, reasoning, and coding. The model, available as open weights under a modified MIT licence, achieves 77.6% on SWE-Bench Verified and 91.4% on τ³-Telecom, and can self-host on four GPUs. Concurrently, Mistral introduced Vibe remote agents for asynchronous cloud coding sessions, launchable from CLI or Le Chat, and a new Work mode in Le Chat (Preview) for multi-step tasks like research and cross-tool actions, both powered by Medium 3.5 for Pro, Team, and Enterprise plans.
Why it matters
This release lowers the hardware barrier for deploying advanced agentic capabilities, enabling platform engineers to run a 128B model on minimal GPU infrastructure. For development teams, the introduction of remote agents and Work mode shifts coding and complex workflows to the cloud, reducing local resource dependency and allowing parallel execution of tasks. This follows Anthropic's launch of Remote Control for Claude Code, but Mistral's open-weight approach provides greater control over model deployment and data. Procurement teams evaluating agentic solutions now have a high-performing, self-hostable option with transparent licensing and API pricing at $1.5/M input and $7.5/M output tokens.
