Coding brokers have principally lived in your laptop computer. As we speak we’re shifting them to the cloud, the place they run on their very own, in parallel, and notify you once they’re executed. You can begin them from the Mistral Vibe CLI or straight in Le Chat, offloading a coding activity with out leaving the dialog.
Powering that is Mistral Medium 3.5 in public preview, our new default mannequin in Mistral Vibe and Le Chat, constructed to run for lengthy stretches on coding and productiveness work. The brand new Work mode in Le Chat (Preview) extends this with a robust agent for advanced, multi-step duties like analysis, evaluation, and cross-tool actions.
Highlights.
- Mistral Medium 3.5, a brand new flagship mannequin that merges instruction-following, reasoning, and coding right into a single 128B dense mannequin. Launched as open weights, below a modified MIT license.
- Sturdy real-world efficiency at a dimension that runs self-hosted on as few as 4 GPUs.
- Mistral Vibe distant brokers for async coding: periods run within the cloud, could be spawned from the CLI or Le Chat, and an area CLI session could be teleported as much as the cloud.
- Begin Mistral Vibe coding duties in Le Chat. Classes run on the identical distant runtime and maintain going when you step away.
- Work mode in Le Chat runs on a brand new agent, powered by Mistral Medium 3.5, that works by multi-step duties, calling instruments in parallel till the job is completed.
Mistral Medium 3.5.
Mistral Medium 3.5 is our first flagship merged mannequin, obtainable in public preview. It’s a dense 128B mannequin with a 256k context window, dealing with instruction-following, reasoning, and coding in a single set of weights. It performs strongly in real-world use, with self-hosting doable on as few as 4 GPUs. Reasoning effort is now configurable per request, so the identical mannequin can reply a fast chat reply or work by a fancy agentic run. We educated the imaginative and prescient encoder from scratch to deal with variable picture sizes and facet ratios.
Mistral Medium 3.5 scores 77.6% on SWE-Bench Verified, forward of Devstral 2 and fashions like Qwen3.5 397B A17B. It additionally has sturdy agentic capabilities and scores 91.4 on τ³-Telecom.
The mannequin was constructed for long-horizon duties, calling a number of instruments reliably, and producing structured output that downstream code can eat. It’s the mannequin that made async cloud brokers in Vibe sensible to ship.
Mistral Medium 3.5 turns into the default mannequin in Le Chat. It additionally replaces Devstral 2 in our coding agent, Vibe CLI.
Vibe distant brokers.
From immediately, coding periods can work by lengthy duties when you’re away. Many can run in parallel, and also you cease being the bottleneck on each step the agent takes.
You can begin the cloud brokers from the Mistral Vibe CLI or from Le Chat. Whereas they run, you possibly can examine what the agent is doing, with file diffs, software calls, progress states, and questions surfaced as you go. Ongoing native CLI periods could be teleported as much as the cloud whenever you wish to depart them working, with session historical past, activity state, and approvals carrying throughout.
Vibe sits between the methods engineering groups already use, with people within the loop wherever they’re wanted. It plugs into GitHub for code and pull requests, Linear and Jira for points, Sentry for incidents, and apps like Slack or Groups for reporting.
Every coding session runs in an remoted sandbox, together with broad edits and installs. When the work is completed, the agent can open a pull request on GitHub and notify you, so that you evaluate the end result as an alternative of each keystroke that produced it.
It matches the high-volume, well-defined work that takes a developer’s time with out taking their judgment: module refactors, check era, dependency upgrades, CI investigations, in addition to bug fixes.
We use Workflows orchestrated in Mistral Studio to deliver Mistral Vibe into Le Chat. We initially constructed this for our personal in-house coding surroundings, then for our enterprise customers. As we speak the potential opens as much as everybody, who can now launch coding duties from the online. And with out being tied to an area terminal, a developer can run a number of in parallel.
You can begin coding periods straight in Le Chat, so a activity described in chat runs on the identical distant runtime because the CLI and the online, and comes again later as a completed department or a draft PR.
New Work mode in Le Chat (Preview).
Work mode is a robust new agentic mode for advanced duties in Le Chat, powered by a brand new harness and Mistral Medium 3.5. The agent turns into the execution backend for the assistant itself, so Le Chat can learn and write, use a number of instruments without delay, and work by multi-step initiatives till it completes what you’ve requested.
Right here’s what Work mode allows you to do immediately.
- Cross-tool workflows: catch up throughout e mail, messages, and calendar in a single run; put together for a gathering with attendee context, newest information, and speaking factors pulled out of your sources.
- Analysis and synthesis: dive into a subject throughout the online, inside docs, and related instruments, then produce a structured temporary or report you possibly can edit earlier than exporting or sending.
- Triage your inbox and draft replies; create points in Jira out of your group and buyer discussions; ship a abstract to your group on Slack.
Classes persist longer than a typical chat reply, so an agent can maintain going throughout many turns, by trial-and-error, and thru to completion. In Work mode, connectors are on by default reasonably than chosen manually, which lets the agent attain into paperwork, mailboxes, calendars, and different methods for the wealthy context it must take appropriate motion.
Each motion the agent takes is seen: you see every software name and the considering rationale. Le Chat will ask for specific approval—based mostly in your permissions—earlier than continuing with delicate duties like sending a message, writing a doc, or modifying knowledge.
Get began.
Mistral Medium 3.5 is obtainable immediately in Mistral Vibe and Le Chat, and powers distant coding brokers and Work mode in Le Chat on the Professional, Crew, and Enterprise plans.
By API, it’s priced at $1.5 per million enter tokens and $7.5 per million output tokens. Open weights are on Hugging Face below a modified MIT license.
It’s also obtainable for prototyping, hosted on NVIDIA GPU-accelerated endpoints on build.nvidia.com and as a scalable containerized inference microservice, NVIDIA NIM.
Construct the way forward for agentic methods with us.
We’re hiring throughout analysis, engineering, and product to push agentic methods additional. See our open roles.