French Mistral this week released Mistral Large 4: 1 trillion parameters (49B activated per inference), trained on a self-built 3,800-GPU Grace Blackwell cluster, scoring 38 on Artificial Analysis—second only to China's DeepSeek 4.1 Flash. We think the more interesting story is that European AI finally has a model worth discussing in the mainstream, but it remains roughly six months behind the global frontier: it took 1 trillion parameters to match what Chinese teams achieved with 552 billion.

What this is

Mistral Large 4 is Mistral France's newest flagship model, released this week under the codename "Le Chonk." It carries 1 trillion total parameters, but activates only 49 billion per inference—a textbook MoE (Mixture of Experts) setup: a large total capacity, but only a small slice of expert networks fires on any given query, which keeps inference costs manageable.

A few key facts:

  • Trained on a self-built cluster of 3,800 NVIDIA Grace Blackwell (GB200) GPUs
  • Available as an API preview only; Mistral has committed to releasing open weights by month-end
  • Artificial Analysis third-party benchmark: 38—second only to DeepSeek 4.1 Flash
  • A leap from Mistral Large 3 (released last December), which scored 9

Industry view

The bullish case: Mistral is "back at the table." Europe had nearly vanished from the first tier of large models; Mistral Large 4 puts it back in the mainstream conversation. One trillion parameters, self-built compute, and month-end open weights—all three legs are in place. A landmark moment for European AI infrastructure.

The bear case / risks: a score of 38 only puts it nose-to-nose with DeepSeek 4.1 Flash—a 552-billion-parameter model. Simon Willison's read: "about 6 months behind the frontier." Caught up, but not caught up. Matching the same performance with more parameters means MoE engineering efficiency still trails China's.

Another underappreciated signal: Mistral's self-built 3,800-GB200 cluster means a European company is no longer fully dependent on US cloud providers and now has end-to-end training capability. This has direct implications for Europe's data-sovereignty regulations going forward.

Impact on regular people

  • For enterprise IT: Once open weights ship at month-end, enterprises can self-host a model near the closed-source frontier without handing all their data to OpenAI or Anthropic.
  • For working professionals: MoE is the new default—"big parameters, cheap inference" is the dominant theme for the next year. Anyone shipping AI products needs to recalculate their cost model.
  • For consumers: Mistral has no consumer product, so end users won't touch this directly. But the existence of a European frontier model puts indirect downward pressure on GPT and Claude API pricing.