Mistral AI unveils Medium 3.5 model and Work Mode for Le Chat

Mistral AI has launched Mistral Medium 3.5, a 128B parameter dense model with a 256k context window, now available in public preview. Developers and organizations can access it via Mistral Vibe CLI, Le Chat, or through the API, targeting Pro, Team, and Enterprise users globally. This release brings coding agents from local machines to the cloud, enabling parallel sessions that notify users upon completion and can be initiated or monitored from web or terminal interfaces. Sessions are isolated in sandboxes with support for broad code edits and can be teleported between local and cloud environments while preserving state and approvals.

Mistral Medium 3.5 integrates instruction-following, reasoning, and coding in a single model, supporting variable image sizes through a newly trained vision encoder. It achieves 77.6% on SWE-Bench Verified and outperforms prior models like Devstral 2 and Qwen3.5 397B A17B in coding and agentic benchmarks.

The model's open weights are released under a modified MIT license, and it is deployable with as few as four GPUs, enabling both self-hosted and cloud-based use cases. Work mode in Le Chat leverages this model to execute multi-step, cross-tool workflows, handling complex research and productivity tasks while surfacing agent actions for user approval. Industry experts highlight the seamless cloud transition and the agent’s capacity for high-volume, well-defined coding work, with early users noting the improved workflow automation and transparency.

Mistral AI, known for its rapid advancements in large language models and agentic tooling for developers, aims to make cloud-based autonomous agents accessible to a broader audience. This move positions the company as a strong competitor in cloud-based AI coding assistance, emphasizing scalability, transparency, and integration with widely used productivity tools.

Source