Mistral claims visual grounding lead on Dense 200

Mistral AI says its Mistral Large 4 model, nicknamed "le Chonk", is strong at visual grounding, meaning it can pinpoint where an object or piece of evidence sits in an image. In its post, Mistral reports 42% on a test called Dense 200 and says that is ahead of frontier closed models. Its launch article gives the specific comparison as 42% versus 41% for GPT-6-Astra.

That is a one-point vendor-reported margin on a single named test, and neither the post nor the article defines Dense 200, so it should be read as Mistral's claim rather than an independently established lead. The article's demonstrations involve technical drawings, PDFs and gigapixel satellite imagery; those are Mistral's own demonstrations, not independent evidence of performance on those workloads. Background on the model is in Intel's coverage of the Mistral Large 4 announcement.