Table of Contents
- What Medium 3.5 Actually Is
- The Open-Weights Decision Still Matters
- Vibe Remote Agents: Cursor on Mistral’s Terms
- Le Chat Work Mode
- How It Stacks Up Against the Competition
- How to Actually Use It
- FAQ
Open weights have been a Mistral signature since Mistral 7B in 2023. Cloud-based asynchronous coding agents have been a Cursor signature for the last year. On May 2, 2026, Mistral shipped both in one release. Mistral Medium 3.5 is a new 128-billion-parameter dense model with a 256K context window, released under a modified MIT license — and the launch shipped alongside Vibe Remote Agents and Le Chat Work Mode, two products that turn the model into an agent platform competing directly with Cursor 3 and Claude Code.
The combination is the move. Open-weights frontier models exist. Cloud-agent platforms exist. Mistral is the first lab to ship both as one coherent stack, and that combination is what makes the release worth paying attention to even by people who do not normally track European AI labs.
What Medium 3.5 Actually Is
Mistral Medium 3.5 is a 128B-parameter dense model with a 256K-token context window. It unifies instruction-following, reasoning, and code generation into a single set of weights — no separate reasoning model, no separate code model. Reasoning effort is configurable per request, which lets the same model handle a quick conversational reply or a deep multi-step agentic run.
On Mistral’s own benchmarks, Medium 3.5 scored 77.6% on SWE-Bench Verified and 91.4% on T3-Telecom. The SWE-Bench number lands in the same range as Grok 4.3 and within shouting distance of Claude Opus 4.7’s 87.6% — close enough for production use, low enough that Mistral cannot claim the frontier crown. The interesting part is that this competitive performance is open weights you can download and run on your own hardware, which the closed-frontier labs do not offer.
The Open-Weights Decision Still Matters
The modified MIT license on Medium 3.5 is the strategic choice. Anthropic, OpenAI, and Google all ship closed weights. Meta ships Llama under a restricted license. Mistral keeps shipping weights you can actually run locally, including in regulated or air-gapped environments where the closed labs cannot deploy.
For most users, that distinction does not matter — they will hit Mistral through Le Chat or the API the same way they hit Claude or ChatGPT. For specific enterprise buyers — defense, healthcare with strict data residency, financial services with sovereign-cloud requirements, governments worried about US frontier-lab dependency — open weights are the deciding factor. Mistral has built a real European business on that specific differentiator, and Medium 3.5 strengthens it without giving up performance.
Vibe Remote Agents: Cursor on Mistral’s Terms
The agent platform half of the launch is Vibe Remote Agents. Coding agents now run in the cloud, in parallel, and notify you when they finish. You can spawn them from the Mistral Vibe CLI or directly inside Le Chat. Each session runs in an isolated sandbox that allows broad edits and installs. When the work is done, the agent can open a pull request on GitHub and notify you.
The teleport feature is the genuinely new mechanic. An ongoing local Vibe CLI session can be lifted up to the cloud mid-task — session history, task state, and pending approvals carry across. The user closes the laptop, goes to dinner, and the agent finishes the migration in the cloud and reports back. Cursor 3 ships cloud agents through its Agents Window, but the local-to-cloud teleport for an in-progress session is Mistral-specific and meaningfully better for the developer who wants to start work locally and walk away from it.
This is the move that makes Mistral interesting as more than a model vendor. The combination of an open-weight model and a first-party cloud agent platform competing on workflow quality, not just weights access, repositions the company in the same agent-stack race that Cursor, Claude Code, and OpenAI Codex are running.
Le Chat Work Mode
The third launch component is Work Mode in Le Chat (Preview), an agent that handles multi-step research, analysis, and cross-tool actions. Le Chat can read and write external data sources, call multiple tools simultaneously, work through long projects, and complete tasks that require sequencing rather than single answers.
The capability list maps to what every other lab’s agent surface offers in 2026 — research with citations, document drafting, structured analysis, action through connected tools. Where Work Mode differentiates is visibility: tool calls and intermediate steps are exposed to the user, and sensitive operations require explicit approval. The honest framing: Le Chat Work Mode is competent but not category-defining. The reason to use it is that you are already on Le Chat or want an agent that runs on a model whose weights you can download.
How It Stacks Up Against the Competition
Against Claude Opus 4.7 on coding: Opus wins on benchmark scores (87.6% vs 77.6% on SWE-Bench Verified) and on production deployment reliability. Medium 3.5 wins on open-weights deployment flexibility and per-token cost for self-hosted workloads.
Against Cursor 3 on agent platform: Cursor wins on IDE-native experience and the depth of its existing developer ecosystem. Vibe Remote Agents wins on the teleport feature and on running against Mistral’s own model rather than depending on partner labs for the model layer.
Against Llama 4 on open weights: Llama 4 Maverick is the comparable open-weight competitor at 400B parameters with a 10M context window. Llama wins on raw scale; Medium 3.5 wins on the bundled agent platform and the license terms (modified MIT versus Llama’s restricted license).
For a developer or enterprise buyer in 2026, the practical question is not whether Mistral leads on any single dimension. It does not. The question is whether the combination of competitive coding performance + open weights + first-party cloud agents fits a specific use case better than the alternatives. For sovereign-cloud, regulated-industry, or self-hosting use cases, the answer is increasingly yes.
How to Actually Use It
Three entry points depending on how you work.
For developers, install the Mistral Vibe CLI, authenticate, and try a single coding task locally to learn the agent’s behavior. Once you trust it for that task class, run a longer session and use the teleport command to move it to the cloud when you step away. The teleport is the feature that justifies the workflow change.
For non-developers who want an agent surface, open Le Chat, switch to Work Mode (Preview), and assign a multi-step research or analysis task that has clear success criteria. The visibility into tool calls is the differentiator — you can see what the agent did and reject specific steps rather than accepting the whole output.
For enterprise buyers evaluating self-hosted deployments, download Medium 3.5 weights and run them through your standard internal evaluation. The modified MIT license permits internal use without commercial restrictions for most cases. Verify license compatibility with your legal team before production deployment.
Skip the Mistral stack if your work is closed-frontier-lab benchmarks-first. Opus 4.7 and GPT-5.5 lead the closed frontier and will keep leading on raw capability until Mistral or another open-weight lab catches up.
FAQ
What is Mistral Medium 3.5?
A 128B-parameter dense model with a 256K-token context window, released May 2, 2026 under a modified MIT license. The model unifies instruction-following, reasoning, and coding into a single set of weights with configurable reasoning effort per request.
How does Medium 3.5 compare to Claude Opus 4.7?
On SWE-Bench Verified, Opus 4.7 scores 87.6% and Medium 3.5 scores 77.6%. Opus leads on raw benchmark performance and production deployment reliability. Medium 3.5 leads on open-weights flexibility, license terms, and self-hosting use cases. The choice depends on whether benchmark-leading capability or deployment flexibility matters more for your work.
What are Vibe Remote Agents?
Cloud-based coding agents that run in isolated sandboxes, in parallel, and notify you when they finish. Spawnable from the Mistral Vibe CLI or directly in Le Chat. The teleport feature lifts an ongoing local CLI session to the cloud mid-task with session history, state, and approvals preserved.
What does Le Chat Work Mode do?
A multi-step agent for research, analysis, and cross-tool actions inside Le Chat. Reads and writes external data sources, calls multiple tools simultaneously, and surfaces tool calls and intermediate steps for user inspection. Sensitive operations require explicit approval. Currently in preview.
Can I download Mistral Medium 3.5 weights?
Yes. The model is released under a modified MIT license that permits internal use without commercial restrictions for most cases. The full license terms should be reviewed by your legal team before production deployment, particularly for revenue-generating use.
Is Mistral catching up to OpenAI and Anthropic?
On raw capability, no — Opus 4.7 and GPT-5.5 still lead the closed frontier. On the combination of competitive capability plus open weights plus first-party cloud agents, Mistral is the only lab shipping all three. That is a different competitive position than “frontier catch-up” and a more defensible one in regulated, sovereign, or self-hosting markets.