Practical Open Weights
The important part of today's Mistral launch is not only the model itself. It is the combination of an open-weights 128B dense model, a 256k context window, and new product surfaces built around longer-running agentic work.
Mistral Medium 3.5 arrives in public preview as a multimodal model optimized for coding and agentic tasks, but the bigger takeaway is strategic. Mistral is tying the model directly to Work mode in Le Chat and remote agents in Vibe, which makes execution, not just conversation, the center of the product story.
Today's News
Mistral introduced Medium 3.5 in public preview in April 2026 as a frontier-class multimodal model for coding and agentic use cases. The official model card lists a 256k context window, modified MIT licensing, and API pricing of $1.5 per million input tokens and $7.5 per million output tokens.
See the Mistral overviewWhy It Matters
Mistral is making Medium 3.5 the default model in Le Chat and the engine behind Work mode and Vibe remote agents. That matters because it shifts attention from single-turn chat performance toward longer-running workflows with tools, approvals, and parallel execution.
Read the model selection guideWhat To Watch
Mistral says Medium 3.5 can be self-hosted on as few as four GPUs and was built for long-horizon tasks with reliable tool use and structured outputs. The next question is whether more teams start treating open models as practical execution backends, not just cheaper chat endpoints.
Read the right-sized AI thesisThat is why Mistral Medium 3.5 matters beyond one benchmark cycle. It shows how the next competitive layer may be the full agent stack: model, tools, approvals, and long-running workflows that teams can actually use in production.
Explore the Mistral family