Mistral Launches Medium 3.5 Model and Cloud-Based Coding Agents
Key Takeaways
- ▸Mistral Medium 3.5 is a 128B open-weight flagship model with 256K context window, achieving 77.6% on SWE-Bench Verified and strong agentic capabilities
- ▸Remote coding agents now execute asynchronously in the cloud via Mistral Vibe, able to run in parallel while keeping developers informed of progress
- ▸Work mode in Le Chat (Preview) enables multi-step task execution with tool-calling and structured output powered by Medium 3.5
Summary
Mistral announced Mistral Medium 3.5, a new 128B flagship open-weight model released under a modified MIT license. Merging instruction-following, reasoning, and coding capabilities into a single dense model with a 256K context window, it achieves 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom—outperforming Devstral 2 and competing with larger models. The model is designed for self-hosting on as few as four GPUs, with configurable reasoning effort per request to balance latency and capability.
Central to the announcement is the introduction of remote coding agents in Mistral Vibe, a paradigm shift that moves agent execution from local laptops to cloud infrastructure. Developers can spawn async coding tasks from the Vibe CLI or Le Chat that run in parallel and notify users upon completion. Local CLI sessions can be "teleported" to the cloud while preserving context and approvals, and each agent runs in an isolated sandbox with integrations to GitHub, Jira, Linear, Sentry, and communication tools like Slack and Teams.
Mistral also previewed Work mode in Le Chat, a new agentic interface powered by Medium 3.5 that handles complex multi-step tasks including research, analysis, and cross-tool orchestration. The model replaces Devstral 2 as the default in both Le Chat and Vibe CLI, reflecting Mistral's confidence in its coding and agentic performance. The architecture positions humans in the loop wherever judgment is needed, targeting high-volume, well-defined work like refactors, test generation, dependency upgrades, and CI investigations.
- Model supports self-hosting on four GPUs, released under modified MIT license, with configurable reasoning effort for flexible performance-latency trade-offs
Editorial Opinion
Mistral's shift of coding agents to cloud infrastructure addresses a real pain point in developer workflows—the need to babysit long-running tasks. The combination of open weights, transparent licensing, and self-hosting viability makes this a compelling alternative to proprietary cloud services. However, the practical impact hinges on reliability, latency, and whether developers trust cloud-based agents with their codebase; early enterprise adoption will be key validation.
