Mistral AI shipped Mistral Medium 3.5 on April 29, 2026 — a 128-billion-parameter open-weight model that merges instruction-following, reasoning, and coding into a single architecture. It launched alongside two new capabilities: Work Mode in Le Chat, an agentic assistant layer for business teams, and remote cloud agents for its Vibe coding environment.

This release matters because it addresses a real bottleneck in enterprise AI adoption: the gap between what AI models can do in a demo and what they can actually execute inside a business workflow.

What Mistral Medium 3.5 Actually Is

Unlike many frontier models that use a Mixture-of-Experts architecture (where only a fraction of parameters activate per inference), Medium 3.5 is a dense model. All 128 billion parameters activate for every token. That design choice has tradeoffs — it’s more compute-intensive per query — but it also means more consistent behavior across instruction types.

The specs:

• 128B dense parameters
• 256,000-token context window
• Self-hostable on as few as four GPUs
• Open weights released under a modified MIT license
• Priced at $1.50 per million input tokens and $7.50 per million output tokens on the API

On SWE-Bench Verified — a benchmark that tests whether a model can generate working patches to fix real GitHub issues — Medium 3.5 scored 77.6%, ahead of competing models including Qwen3.5 397B A17B. On τ³-Telecom, which tests agentic tool use in specialized environments, it scored 91.4%.

Configurable reasoning effort is built in, so the same model can handle a quick lookup or a complex multi-hour agentic run depending on how you call it.

Work Mode in Le Chat: The Business-Relevant Part

The model is interesting. The Work Mode feature is where things get practically relevant for enterprise teams.

Work Mode turns Le Chat — Mistral’s consumer and enterprise assistant — into a proper autonomous agent that can complete multi-step tasks across multiple tools simultaneously. Instead of answering questions, it acts on them.

What it can do:

• Connect to email and calendar accounts by default
• Execute parallel tool calls across Jira, Slack, documents, and data sources
• Work through complex tasks in the background without requiring constant oversight
• Surface every tool call and reasoning step so you can see exactly what it did
• Require explicit human approval before taking sensitive actions (sending emails, modifying records)

That last point is the trust mechanism most enterprise AI implementations are still missing. Letting AI agents act autonomously is easy. Letting them act autonomously while maintaining a clear audit trail and requiring approval for irreversible actions is the design that actually works in regulated or customer-facing environments.

Work Mode is available on Le Chat Pro, Team, and Enterprise plans.

Remote Agents in Vibe: Async Cloud Coding

The third piece of this release is aimed at developers and technical teams. Vibe, Mistral’s coding environment, now supports remote agents that run asynchronously in the cloud.

The practical effect: you start a coding task, the agent picks it up and runs it in the cloud, and you walk away. You can check back later, review diffs and tool calls, and approve or redirect the work. Long-running local CLI sessions can also be teleported to the cloud mid-session, with full context preserved.

This shifts AI-assisted development from a synchronous, back-and-forth model to one that resembles delegating work to a capable colleague who gives you updates and asks for sign-off before making significant changes.

What This Means for Business

Open weights with enterprise-grade performance is a real change. For years, open-weight models lagged behind closed frontier models on real-world tasks. Mistral Medium 3.5’s benchmark scores suggest that gap is narrowing in ways that matter. For businesses that want to run AI on their own infrastructure — for data privacy, cost control, or compliance reasons — that matters.

Agentic work assistants are arriving in productivity tools, not just developer environments. Work Mode in Le Chat is not aimed at engineers. It’s aimed at knowledge workers: people who coordinate across email, calendars, documents, and project tools. As this capability shows up in more business software (Salesforce, Microsoft, Google are all shipping versions of this), the ability to configure and govern these agents becomes a core business competency, not an IT project.

Audit trails and approval gates are the enterprise differentiator. The fact that Mistral built explicit approval requirements and full tool call visibility into Work Mode is not accidental. Enterprise buyers won’t deploy autonomous AI agents without them. If your AI vendor can’t show you what the agent did and why, that’s a hard block for any compliance-aware business.

Self-hosting is back on the table. At 128B parameters self-hostable on four GPUs, Medium 3.5 is the kind of model that a mid-sized company with a real data infrastructure could run internally. That option — running frontier-level AI on your own compute — was effectively closed for the last two years. It’s reopening.

The EDNA Take

We track these releases closely because they change what’s possible for the businesses we work with. A few months ago, giving knowledge workers a capable autonomous assistant that could act across their toolstack required custom development and significant infrastructure. Work Mode and similar releases from OpenAI and Google are moving that into standard enterprise software.

The question isn’t whether to use agentic AI in your business. The question is how to configure, govern, and get value from it without creating new operational risks. That’s where EDNA’s work on AI strategy and implementation comes in — helping teams understand not just what these tools can do, but what the right deployment looks like for their specific workflows and risk profile.

The practical next step is the free Working With Claude field guide. Thirty-two pages covering the ecosystem, Claude Code, and how to govern a rollout properly. Get your copy.