Mistral coding agents leave your laptop behind: parallel sessions, no hovering required

Mistral has pulled coding agents off the local terminal and into its cloud. Developers can now fire off long-running tasks and check back later, instead of babysitting a command line. The update, announced today, rests on Mistral Medium 3.5, a dense 128B parameter model that wraps instruction-following, reasoning, and coding into a single weight set, and introduces remote agents in both the Mistral Vibe CLI and Le Chat, plus a new Work mode for multi-step workflows. DeepSeek's DSpark just fixed the two things that held…

Mistral frames the shift as a fix for a structural problem: coding agents have mostly been stuck on local machines, forcing developers to hunker down at every step. With remote agents, sessions run in isolated cloud sandboxes. You can start them from the command line or a web chat, and the agent pings you when it's done. Multiple agents can run in parallel, and if you have a local session going, you can teleport it to the cloud with the full session history intact. Microsoft's new platform gives scientists a governed…

Mistral Medium 3.5: a single model for long-horizon tasks

Mistral Medium 3.5 is the company's first flagship merged model, released as open weights under a modified MIT license. It scores 77.6% on SWE-Bench Verified, ahead of Devstral 2 and Qwen3.5 397B A17B. The model also hits 91.4 on τ³-Telecom, a telecom domain benchmark. It is dense, not a mixture of experts, and runs on as few as four GPUs. Mistral kept that deliberate, to keep self-hosting viable for smaller engineering teams. MiniMax launches M2.7 model with strong software…

The model handles a 256k context window and supports configurable reasoning effort per request. That means the same weights can serve a quick chat reply or an extended agentic run. Mistral also trained a vision encoder from scratch to handle variable image sizes and aspect ratios, though the company did not push vision as a primary use case for the cloud agent launch.

Vibe remote agents: parallel coding in the cloud

Mistral Vibe, the coding agent CLI, now supports remote agents. A developer starts a session from the terminal or Le Chat, and the session runs in a cloud sandbox while the developer inspects file diffs, tool calls, and progress states. The agent can be told to open a pull request on GitHub when it's done, notifying the developer for review rather than demanding approval on every keystroke. IBM's open-source CUGA agent harness skips the…

The system integrates with GitHub for code and pull requests, Linear and Jira for issues, Sentry for incidents, and Slack or Teams for reporting. Each session runs in an isolated sandbox, which Mistral says enables broad edits and package installs without contaminating other work. The company targets high-volume, well-defined tasks: module refactors, test generation, dependency upgrades, CI investigations, and bug fixes.

Work mode in Le Chat: multi-step agents with parallel tool calls

Le Chat now includes a Work mode, powered by Mistral Medium 3.5 and a new agent harness. In Work mode, the assistant uses multiple tools concurrently, email, calendar, web search, internal documents, to complete multi-step tasks. The agent persists across many turns and can trial-and-error its way to completion. Mistral says connectors are turned on by default in this mode, so the agent can reach into documents, mailboxes, and calendars without manual selection.

Work mode currently supports cross-tool workflows like preparing for a meeting with attendee context and latest news, research and synthesis into structured briefs, inbox triage with draft replies, and creating Jira issues from team discussions with Slack summaries. Every tool call and reasoning step is visible to the user, and Le Chat asks for explicit approval before sending messages, writing documents, or modifying data. Anthropic Unveils Claude Sonnet 5, Claude Science, and…

Mistral Vibe and Work mode in Le Chat are available on Pro, Team, and Enterprise plans. Mistral Medium 3.5 is priced at $1.5 per million input tokens and $7.5 per million output tokens through the API. The open weights are on Hugging Face, and the model is also hosted on NVIDIA GPU endpoints on build.nvidia.com and as an NVIDIA NIM microservice.

Mistral is hiring across research, engineering, and product roles to keep developing agentic systems.