Mistral Medium 3.5 Launches: 128B Model Powers New Agentic Work Mode
3 min readMistral AI has released Medium 3.5, a dense 128-billion-parameter model that now powers the company’s Le Chat assistant and brings a new agentic “Work mode” capable of running multi-step tasks across email, calendar, documents, and more.
What Is Mistral Medium 3.5?
Medium 3.5 is a dense 128B model with a 256,000-token context window. Unlike mixture-of-experts architectures that activate only a subset of parameters per inference, Medium 3.5 runs all 128 billion parameters on every request. The model handles instruction-following, reasoning, and code generation in a single set of weights, and its vision encoder was trained from scratch to handle variable image sizes and aspect ratios.
On SWE-Bench Verified, a widely used benchmark for evaluating software engineering capability, Medium 3.5 scores 77.6%. That puts it ahead of Devstral 2 and larger sparse models like Qwen3.5 397B A17B, according to Mistral’s official announcement.
Work Mode: Agentic Tasks Come to Le Chat
The more immediately practical addition is Work mode in Le Chat, Mistral’s consumer-facing AI assistant. Work mode introduces a multi-step agentic layer: the model can call tools in parallel, reason through complex tasks, and work across integrations including email, calendar, Jira, Slack, and documents. Every tool call and reasoning step is visible to the user, and the system requires explicit approval before taking sensitive actions.
Mistral is positioning this as a direct competitor to OpenAI’s operator-style agents and Microsoft Copilot’s enterprise workflows. Work mode is available to Pro, Team, and Enterprise subscribers.
Remote Agents in Vibe
Alongside Le Chat, Mistral also launched remote agents in Vibe, its developer coding environment. Remote agents run asynchronous cloud-based coding sessions, meaning developers can kick off a complex refactor or test-writing task and let the model run in the background without keeping a local session open. This mirrors similar async capabilities recently introduced by competitors in the coding-assistant space.
Why It Matters
The 77.6% SWE-Bench score is notable because it positions a European AI lab at the front of the coding benchmark leaderboard with a publicly accessible model. For enterprise users, the Work mode integration means Mistral is no longer just a model provider but an end-to-end workflow automation platform with real-time transparency into what the agent is doing.
The launch also intensifies competition in the 100B-plus parameter segment of the open-weights and commercial model market. With pricing drawing some criticism from early users, Mistral’s ability to retain developers will likely hinge on whether the performance advantage holds up in production workloads over the coming weeks.
What to Watch Next
Mistral has not announced a release date for a Large or Ultra tier model in the same series. Given the pace of the company’s releases in 2026, a follow-up announcement is plausible within the next few months. Developers integrating Medium 3.5 via the Mistral API should note that Work mode is currently exclusive to Le Chat’s consumer and enterprise products.