Mistral has launched its new flagship model, Mistral Medium 3.5, combining chat, reasoning, and code functions into a single dense model with 128 billion parameters and a 256,000-token context window. Unlike many competitors that use Mixture of Experts (MoE) architectures to reduce inference costs, Mistral opted for a dense design, which activates all parameters for every token. This trade-off prioritizes simplicity and production stability over cost efficiency, though the company notes the model can be self-hosted on four GPUs.
The model introduces a reasoning_effort parameter that lets users toggle between quick responses and deeper reasoning for complex tasks. Mistral also built a new vision encoder from scratch to handle variable image sizes and aspect ratios. In benchmarks, Medium 3.5 scored 77.6% on SWE-Bench Verified and 91.4% on T3-Telecom, but lagged behind Claude in banking scenarios.
Weights are available on Hugging Face under a Modified MIT License, which allows commercial and non-commercial use but restricts high-revenue companies. API pricing is set at $1.50 per million input tokens and $7.50 per million output tokens.
Beyond the model, Mistral enhanced its coding tool Vibe with asynchronous cloud agents that run in isolated sandboxes and can open pull requests. The agents integrate with GitHub, Linear, Jira, Sentry, Slack, and Teams. Additionally, Le Chat gains a Work Mode that enables multi-step tasks across tools like email, calendars, and documents, with explicit user confirmation for sensitive actions. Work Mode is available on Pro, Team, and Enterprise plans.