Mistral AI Unveils Medium 3.5 and Vibe Remote Agents
Mistral AI announced Mistral Medium 3.5, remote coding agents in Vibe, and a new Work mode in Le Chat for tackling complex tasks. These updates shift coding agents from local laptops to the cloud. There, they operate independently, handle multiple sessions at once, and alert users upon completion. Users can launch them via the Mistral Vibe CLI or within Le Chat conversations, passing off coding work seamlessly.
Mistral Medium 3.5 drives these features. It entered public preview as the new default model in Mistral Vibe and Le Chat. This model supports extended runs on coding and productivity jobs. Le Chat's Work mode builds on it with an agent designed for intricate, multi-step processes such as research, analysis, and actions across tools.
Mistral Medium 3.5 Details
Mistral Medium 3.5 marks the company's first flagship merged model. It combines instruction-following, reasoning, and coding in one 128B dense setup. Available as open weights under a modified MIT license, it delivers solid performance in practical settings. Teams can self-host it on just four GPUs.
The model features a 256k context window. Users can adjust reasoning effort per request, allowing quick responses or deep agentic processing. Mistral AI trained its vision encoder from the ground up to manage varying image sizes and aspect ratios.
On benchmarks, it achieves 77.6% on SWE-Bench Verified, surpassing Devstral 2 and models like Qwen3.5 397B A17B. It excels in agentic tasks with a 91.4 score on τ³-Telecom. Designed for long tasks, it calls tools repeatedly and generates structured outputs for code integration. This capability enabled practical async cloud agents in Vibe.
Mistral Medium 3.5 now serves as the default in Le Chat and replaces Devstral 2 in the Vibe CLI coding agent. Access comes through API at $1.5 per million input tokens and $7.5 per million output tokens. Open weights appear on Hugging Face. It also runs on NVIDIA GPU-accelerated endpoints at build.nvidia.com and as a containerized inference service via NVIDIA NIM.
Mistral AI, a French startup founded in 2023 by former DeepMind researchers, specializes in efficient open-source large language models. Early successes include the Mistral 7B model, which gained attention for outperforming larger rivals despite its size.
Vibe Remote Agents
Vibe remote agents let coding sessions tackle lengthy tasks without user supervision. Multiple agents run concurrently, removing delays from sequential steps. Launch them from the Mistral Vibe CLI or Le Chat. While active, users view file changes, tool calls, progress, and queries in real time.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Local CLI sessions transfer to the cloud easily, preserving history, state, and approvals. Vibe connects to tools developers use, including GitHub for code and pull requests, Linear and Jira for issues, Sentry for incidents, and Slack or Teams for updates.
Each session operates in an isolated sandbox that permits extensive edits and installations. Upon finishing, agents create GitHub pull requests and send notifications, letting users review outcomes rather than individual actions.
These agents suit repetitive tasks like module refactors, test creation, dependency updates, CI checks, and bug fixes. Workflows from Mistral Studio integrate Vibe into Le Chat. Originally for internal and enterprise use, this now reaches all users. Developers start tasks from the web, running several in parallel without a local terminal.
Sessions in Le Chat use the same remote runtime as CLI and web versions. They return as completed branches or draft PRs.
Work Mode in Le Chat
Work mode, in preview, equips Le Chat with a strong agent powered by Mistral Medium 3.5 and a new framework. The agent handles execution, enabling reads, writes, parallel tool use, and multi-step projects to completion.
Current capabilities cover cross-tool workflows, such as syncing email, messages, and calendars; or prepping meetings with attendee details, news, and points from sources. It supports research across web, docs, and tools, yielding editable briefs or reports.
Users triage inboxes, draft replies, create Jira issues from discussions, or send Slack summaries. Sessions last beyond single replies, persisting through iterations to finish. Connectors activate by default for context from documents, mailboxes, calendars, and systems.
All agent actions show clearly: tool calls and rationales appear. Le Chat seeks explicit approval for sensitive steps like messaging, document edits, or data changes, respecting permissions.
Mistral Medium 3.5, remote agents, and Work mode require Pro, Team, or Enterprise plans in Vibe and Le Chat.
Related on Neura Market