Mistral AI Launches Medium 3.5 Model with Cloud Agents and Work Mode

The French AI company Mistral AI has officially released Mistral Medium 3.5, a 128B parameter dense model now available in public preview. This announcement marks a significant shift in how developers interact with AI coding assistants, moving sessions from local machines to cloud-based execution environments.

According to the official announcement from Mistral, the new model consolidates instruction-following, reasoning, and coding capabilities into a single set of weights. It features a 256k context window and can be self-hosted on as few as four GPUs, making it accessible for both cloud and on-premises deployments.

The core innovation here isn't just the model itself. It's the infrastructure built around it. Mistral's official blog post details how coding agents now run in isolated cloud sandboxes, handling long-running tasks while developers step away. You can spawn these sessions from the Mistral Vibe CLI or directly within Le Chat, then receive notifications when work completes.

This changes the physical workflow. Instead of watching a terminal cursor blink through hours of dependency upgrades or test generation, you initiate the task and return to find a pull request waiting for review. The agent handles the keystrokes; you handle the judgment calls. Sessions can even be teleported from local CLI environments to the cloud, preserving state and approval history across the transition.

Performance metrics matter here. Mistral Medium 3.5 scores 77.6% on SWE-Bench Verified, outperforming prior models like Devstral 2 and competing with larger architectures including Qwen3.5 397B A17B. The company also reports a 91.4 score on τ³-Telecom, indicating strong agentic capabilities for multi-step workflows.

What's more interesting than the benchmarks is the licensing. The model weights are released under a modified MIT license, which is unusually permissive for a flagship model of this capability tier. This means organizations can deploy it without the restrictive terms that typically accompany frontier AI systems. (A problem that has plagued enterprise adoption for years, frankly.)

The Work Mode feature in Le Chat extends this capability beyond coding. It's a new agentic mode that handles complex, multi-step tasks across connected tools. The agent automatically selects connectors and executes parallel tool calls when dependencies allow. You see each action stream in real time, with explicit approval prompts before sensitive operations like sending emails or modifying records.

Documentation from Mistral shows Work Mode supports cross-tool workflows, research synthesis, inbox triage, and recurring deliverables. The agent pulls from email, calendars, internal docs, and third-party systems like GitHub, Linear, Jira, Slack, and Teams. Output surfaces on Canvas, where users can preview and edit briefs, reports, and slide outlines before finalizing.

There are limitations worth noting. The preview version locks the mode after a conversation starts—you cannot switch back to Fast or Think modes mid-session. Short queries may take longer than in standard chat modes. Some users have reported false tool-access errors requiring retries. These are preview-stage friction points, but they matter for production workflows.

Access is rolling out across Free, Pro, and Team plans, though capabilities and limits may change before general availability. The model replaces Devstral 2 as the default in both Le Chat and the Vibe CLI, which means existing installations need updates to version 2.9.1 or later to recognize the new model alias.

Industry observers note this positions Mistral as a direct competitor to cloud-based AI coding assistance from major players. The emphasis on transparency—visible tool calls, approval gates, session history—addresses a key concern in enterprise AI adoption. Developers want to know what the agent is doing, not just trust a black box.

The pricing structure has shifted alongside the release. Medium 3.5 costs $1.5 per million tokens, while Mistral Large sits at $0.5 per million. This inversion suggests Medium 3.5 may serve as the new flagship for many use cases, with Large reserved for specific high-complexity scenarios. Whether users actually pay for it remains the real question.

Early testing reveals the model handles high-volume, well-defined work effectively: module refactors, test generation, dependency upgrades, CI investigations, and bug fixes. It's less about replacing developer judgment and more about automating the repetitive scaffolding that consumes time without requiring expertise.

For organizations evaluating this, the self-hosting capability is significant. Four GPUs can run the model, which means smaller teams can deploy it without massive infrastructure commitments. The EAGLE head for speculative decoding further improves throughput for production deployments.

Whether this becomes the standard for cloud-based AI coding assistance depends on reliability at scale. The preview period will reveal how well the system handles concurrent sessions, error recovery, and integration edge cases. Time will tell if the promises match the performance.

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn