Mistral AI just dropped something that should make the big US labs nervous. Medium 3.5 isn’t just another model — it’s a 128B parameter flagship that ships with open weights, runs remote coding agents in the cloud, and costs nothing to evaluate.
The Paris-based company announced the release on April 29, positioning it squarely against the agentic AI wave that’s dominated 2026. The model hits 77.6% on SWE-Bench Verified, ahead of Devstral 2 and Qwen3.5 397B. Not bad for something you can self-host on four GPUs.
Remote Agents: Your Laptop Can Stop Being the Bottleneck
The headline feature is Mistral Vibe’s remote agents. Instead of tying up your local machine for hours, you can now spawn coding sessions that run in Mistral’s cloud, work through long tasks autonomously, and notify you when they’re done. Multiple agents can run in parallel.
You can start them from the Vibe CLI or directly in Le Chat. Local sessions can even be “teleported” to the cloud — your task state, session history, and approvals carry across seamlessly.
This is the model that made async cloud agents practical to ship. Built for long-horizon tasks, structured output, and reliable multi-tool calls.
Open Weights Under Modified MIT Licence
Here’s where it gets interesting for NZ businesses watching their AI budgets. Medium 3.5 ships as open weights under a modified MIT licence. You can self-host it on as few as four GPUs. For companies wanting to keep their AI workloads on-premises — whether for data sovereignty, cost control, or latency — this is a serious option.
With a 256k context window and configurable reasoning effort per request, the same model handles quick chat replies and complex agentic runs. The vision encoder was trained from scratch for variable image sizes.
Why This Matters
The agentic AI race isn’t just OpenAI versus Anthropic anymore. Mistral’s move shows European AI players aren’t just keeping pace — they’re shipping features the US giants haven’t matched yet. Remote cloud agents with human-in-the-loop approvals, integrated with GitHub, Linear, Jira, and Sentry? That’s a dev workflow, not a demo.
For NZ businesses evaluating multi-vendor AI strategies, having an open-weight, self-hostable option that’s actually competitive changes the calculus. You don’t have to choose between capability and control anymore. Mistral is betting you can have both.
The question isn’t whether European AI can compete. It’s whether the incumbents can match this level of openness without giving up their moat.
🔍 THE BOTTOM LINE
Mistral Medium 3.5 is the most credible open-weight challenger to the proprietary agent platforms yet. Remote cloud coding agents, self-hostable on four GPUs, competitive benchmarks — this isn’t a research paper, it’s a product. If you’re a NZ business that’s been waiting for a viable on-prem AI option, your wait might be over.
