Mistral Large 4 preview hits 38 on Artificial Analysis

Mistral's own cluster of 3,800 NVIDIA Grace Blackwell GPUs marks a shift from the company's earlier reliance on outside compute, and it trained a 1 trillion parameter model with 49 billion active parameters on that hardware. The API preview supports only two reasoning levels — "none" and "high" — and the article notes that the "high" pelican looked better even though it used fewer output tokens, 2,717 versus 3,275 for "none."

The score of 38 on Artificial Analysis is a huge improvement on last December's Mistral Large 3, which scored 9 and drew what the article calls a terrible pelican. It places Mistral Large 4 just behind DeepSeek 4.1 Flash, a 552B model, on that benchmark. The article's own read is that Mistral is back to being maybe about six months behind the frontier, though it is not a Fable-class model.

The immediate question hanging over the release is whether the promised open weights actually arrive at the "end of this month," and whether a shipped open version holds the preview's score. For teams that prefer to self-host, that timing is the difference between a benchmark result and a model they can actually run.