Mistral's Le Chonk Beats GPT-6-Astra in Image Test

French AI startup Mistral has introduced Mistral Large 4, natively multimodal software nicknamed Le Chonk, featuring advanced visual grounding abilities. On the Dense 200 benchmark, the model achieved a 42% score, narrowly topping the 41% posted by OpenAI's closed GPT-6-Astra. Built as a 1 trillion parameter mixture-of-experts architecture with 49 billion active parameters, the flagship system translates visual prompts into precise location coordinates. Developers can now access the model through the Mistral Studio API at $1.36 per million input tokens and $4.18 per million output tokens, with full open weights expected by the end of the month following security red-teaming.

The system excels at dense visual analysis, showing high proficiency on tests like ChartQA Pro alongside its grounding benchmarks. Outside of vision, independent evaluations placed it 32nd among 44 models on a broad benchmark suite, though it demonstrated strong results in domain-specific legal and cybersecurity assessments. Unlike open offerings from Chinese competitors like DeepSeek, Moonshot AI's Kimi, and Alibaba's Qwen, Mistral promises full open-weight downloads for local deployment.

For software engineers and data practitioners, visual grounding transforms how vision-language models integrate into enterprise workflows. Instead of generating descriptive text, Le Chonk produces machine-readable bounding boxes—frequently normalized to a 0-to-1000 range—that point directly to features in engineering drawings, scanned PDF documents, and gigapixel satellite images. These direct spatial outputs allow downstream automation systems to trigger targeted component inspections, extract exact evidence, or execute localized geographic searches on private infrastructure without routing sensitive files to third-party clouds.