Mistral AI announces multimodal model Mistral Large 4, achieving 42% on visual grounding benchmark

Mistral AI has released its new 1-trillion-parameter multimodal model, Mistral Large 4. The company states the model achieved a score of 42% on the Dense200 visual grounding benchmark, narrowly outperforming GPT-6 Astra's 41%. This benchmark tests a model's ability to locate specific objects within an image, not just recognize its general content. Mistral positions this capability as a core feature for applications in fields like cybersecurity, finance, and law. Businesses are advised to treat the benchmark result as an indicator of potential, not a guarantee of performance in all specific use cases.