French AI company Mistral has introduced Mistral Large 4, its new flagship model internally nicknamed "Le Chonk." The model uses a mixture-of-experts, or MoE, architecture with 1.05 trillion total parameters, while activating 49 billion parameters per inference. It natively supports multimodal image-and-text input and offers a context window of up to 1 million tokens. Mistral said the API preview is already live, and the model weights are scheduled to be released on Oct. 27.
Mistral described Large 4 as the strongest open-weight model in overall performance among models developed in the U.S. and Europe. In benchmarks published by the company, the model scored 62% on the DeepSWE v1.1 software engineering test, compared with 61% for GLM-5.3, 57% for DeepSeek V4 Pro, and 51% for Qwen 3.8 Max. It also posted 67% on the Finch finance benchmark, matching DeepSeek V4 Pro, and 73% on the DIOR-RSVG satellite image target localization task, above GPT-6 Astra's 68%.
Those figures, however, were mainly self-reported by Mistral rather than produced under a unified testing setup. Zhipu previously reported a 66.9% score for GLM-5.3 on the same DeepSWE v1.1 benchmark, while Kimi K3 reached 67.5%, both above Large 4's 62%. Mistral said Large 4 was trained from scratch over roughly two months using about 4,000 Nvidia Grace Blackwell GPUs in its own European data centers, while also pitching European infrastructure sovereignty as a key selling point.
French AI company Mistral has launched Mistral Large 4, its new flagship model, internally nicknamed 「Le Chonk」. The model uses a mixture-of-experts architecture with 1.05 trillion total parameters, while only 49 billion parameters are activated during each inference.
Large 4 natively supports image-and-text input and offers a context window of up to 1 million tokens. Mistral has opened an API preview and plans to release the model weights on Oct. 27.
Benchmark results published by Mistral
Mistral said Large 4 is currently the strongest open-weight model in overall performance among models developed in the United States and Europe. In the company's published results, Large 4 scored 62% on the DeepSWE v1.1 software engineering benchmark. In Mistral's comparison table, GLM-5.3 scored 61%, DeepSeek V4 Pro scored 57%, and Qwen 3.8 Max scored 51%.
On the Finch finance benchmark, Large 4 posted 67%, matching DeepSeek V4 Pro. On the DIOR-RSVG satellite image target localization benchmark, it scored 73%, above GPT-6 Astra's 68%.
Results are not based on a unified testing setup
Those results are currently drawn mainly from Mistral's own testing rather than a unified evaluation standard. When Zhipu released GLM-5.3, it reported a 66.9% score on the same DeepSWE v1.1 benchmark. Kimi K3 reached 67.5%, also above Large 4's 62%. Different teams may have used different test configurations and runtime setups.
Training scale and European infrastructure pitch
Mistral said Large 4 was trained from scratch. The company used about 4,000 Nvidia Grace Blackwell GPUs and trained the model for roughly two months in its own data centers in Europe.
Mistral is also pitching European autonomy as part of the product's appeal. According to the company, model training, API services, and future self-hosted deployment can all remain on European infrastructure.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
