Mistral Medium 3.5 Launch: 128B Dense Model with 256k Context and Cloud Coding
Open-weight model runs on 4 GPUs, powers async cloud coding agents in Le Chat.
Mistral AI introduced Mistral Medium 3.5, a 128B dense flagship model available in public preview. The model features a 256k context window and unifies instruction-following, reasoning, and coding into a single set of weights. It achieves 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom for agentic capabilities. The model is released as open weights under a modified MIT license and can be self-hosted on as few as four GPUs. Reasoning effort is configurable per request, allowing the same model to handle both quick chats and complex agentic runs.
Alongside the model, Mistral launched remote Vibe coding agents that run asynchronously in the cloud. Developers can start sessions from the CLI or directly in Le Chat, with sessions persisting even when the user steps away. The agents can inspect file diffs, tool calls, and progress, and can be teleported from local CLI to the cloud. Vibe integrates with GitHub, Linear, Jira, Sentry, and Slack/Teams, and each session runs in an isolated sandbox. Le Chat also introduces Work mode for multi-step research and analysis tasks, powered by Medium 3.5.
- Mistral Medium 3.5 is a dense 128B model with 256k context, scoring 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom.
- Remote Vibe coding agents run asynchronously in the cloud, can be started from CLI or Le Chat, and support parallel sessions.
- The model is released as open weights under modified MIT license, self-hostable on as few as four GPUs.
Why It Matters
Mistral's open-weight, GPU-efficient model and cloud coding agents promise to democratize advanced AI for real-world development workflows.