Mistral has released Mistral Medium 3.5, a 128-billion parameter model designed to handle instruction following, reasoning, and coding within a single system, and introduced new cloud-based agent capabilities in its Vibe and Le Chat products. The model is available in public preview with open weights under a modified MIT license and supports a context window of up to 256k tokens. It can be self-hosted on a small number of GPUs and allows configurable reasoning effort per request, enabling both short responses and longer multi-step executions.
The release also introduces remote coding agents in Mistral Vibe, shifting execution from local environments to cloud-based runtimes. Developers can start coding sessions from a command-line interface or from within Le Chat, where tasks continue running asynchronously. Sessions can be moved from local execution to the cloud, preserving state and history, and multiple agents can run in parallel. Each session runs in an isolated environment, where the agent can modify code, install dependencies, and interact with external systems. When tasks are completed, agents can generate outputs such as pull requests and notify users for review.
Mistral Medium 3.5 is used as the default model for these agents and replaces earlier models in the Vibe CLI. It is designed for long-running workflows, including multi-step tasks that require tool usage and structured outputs. The model also includes a vision encoder trained to handle variable image inputs.
In addition, Mistral has introduced a new Work Mode in Le Chat, which enables an agent to execute multi-step workflows across connected tools. In this mode, the system can access external data sources, perform analysis, and take actions such as drafting messages, creating issues, or generating reports. The agent operates with visibility into its actions, including tool calls and intermediate steps, and requires user approval for sensitive operations. Sessions persist across multiple steps, allowing the agent to iterate until a task is completed.
Community reaction in X has been largely positive, with praise for the seamless local-to-cloud handoff, excitement about EU AI momentum, and appreciation for a capable dense model that runs on fewer GPUs.
Developer Jarek Sobiecki posted:
New model - So far, so good . A noticeable improvement over DevStral 2! So far, I have tested that it works with Helm templates, improvements on GitLab pipeline or creating end-to-end tests. It aligns well with expectations and shows no random quirks. This is really good work!
Other users are already planning to switch back to Vibe, while others note pricing comparisons and early performance tests like user @gioelerosana who commented:
Guys. 1.5$ In / 7.5$ Out it's too expensive for its size. (Gemini 3 Flash is 0.5$ in / 3$ out)
Compared to tools like OpenAI Codex, Cursor, and Claude Code, Mistral focuses on a developer-oriented approach with open weights, self-hosting options, and cloud-based agent execution. Its remote agents also reflect a broader shift toward asynchronous, multi-step AI workflows that run independently and integrate into development pipelines.
The system integrates with developer tools such as GitHub, Jira, and Slack, allowing agents to operate within existing workflows. The release reflects a shift toward running AI agents as asynchronous services in the cloud, with orchestration handled by a combination of model capabilities and external tool integrations.