Mistral Large 4 Released: European AI Tries to Catch Up with Chinese Open Models

What Was Released

Mistral AI released its flagship Mistral Large 4 on October 6. It’s the company’s first major version since December 2025 — the last update was back in April 2026. Co-founder Guillaume Lample said the new model “narrows the gap with the strongest models” and ranks “among the best open models in the world” on aggregate leaderboards.

So they say. These are Mistral’s own reported numbers; independent third-party rankings haven’t confirmed them yet — keep that in mind.

Report Card: Matching GLM 5.3 on Long Coding

Mistral shared preliminary data: on the Deep SWE 1.1 long-coding benchmark, Large 4 scored 63%, on par with Z.AI’s GLM 5.3, ranking 12th overall. The best models today reach 74%.

It also claims parity with the best open models on several niche tasks: finance (e.g., processing large data tables), cybersecurity, geospatial analysis (identifying objects in satellite imagery, assessing natural disaster damage). The company wants to push into industrial design and manufacturing too — Airbus and BMW are already customers.

CEO Arthur Mensch declared at Abu Dhabi’s AI Everything summit that the new model beats Chinese models in specific areas like cybersecurity, and that “Europe can’t compete” doesn’t hold. He didn’t say which Chinese models or what benchmarks, though.

Trained on Just 4,000 Chips

Lample revealed that Large 4 was trained on only 4,000 top-tier chips — roughly a third to half of what Chinese startups use. At the US giants, the starting point is “hundreds of thousands.”

Mistral’s playbook is clear: can’t win on compute, win on efficiency. The company is also building its own data centers in Europe, announcing major investment this summer with a target of 1 gigawatt of capacity by 2030 — half for its own training, half for customers running AI.

Open Weights on October 27

Large 4 is going the open route, with Mistral promising full open weights on October 27. Until then it’s API-only, with the company monitoring for abuse and finishing safety testing. With recent AI model incidents, that step can’t be skipped.

After open-sourcing, it becomes the default model for the Mistral assistant Vibe. Lample added that Vibe also offers Chinese models like GLM — which remain stronger in some areas.

The Europe Card: Sovereignty Narrative

Mistral is playing the “European sovereignty” card: enterprises can download, fine-tune, and deploy on their own servers. Chief scientist Pierre Stock said: “Open AI also guarantees that access to models can never be cut off.” He was alluding to the White House’s temporary restrictions on Anthropic’s Mythos and Fable.

How Far Behind

Mistral admits the gap with top US and Chinese models persists. It had previously dropped to 24th on Artificial Analysis’s aggregate ranking, behind OpenAI, Anthropic, Google, and Chinese companies like Moonshot, Alibaba, and Z.AI. The company says Large 5 and Large 6 will keep climbing the same technical path next year.

Reserve judgment until independent evaluations land.