Mistral says latest frontier model is most capable outside US and China

French AI company Mistral has announced Mistral Large 4 (ML4), its largest and most sophisticated model yet, is a competitor to leading models across text output, coding and visual reasoning.

ML4, nicknamed ‘le Chonk’, is an open-weight, natively multimodal mixture-of-experts model with one trillion parameters and 49 billion active parameters. It was designed to support more than 160 languages.

Mistral said that in the coding benchmark DeepSWE 1.1, which measures the performance of coding agents across extended coding tasks, ML4 achieved 62 per cent. This puts it ahead of leading Chinese AI models such as Alibaba’s Qwen 3.8 Max, DeepSeek’s V4 Pro and Z.ai’s GLM-5.3.

The independent AI benchmarking platform Artificial Analysis said ML4 will become one of the top three open-weight models on its cyber index, and compared its overall performance to that of OpenAI’s GPT-6 Sol.

In other areas, such as visual reasoning, Mistral said ML4 outperforms the most powerful AI models on the market. It said that in Dense200, a benchmark that tests object detection capabilities, ML4 scored 42 per cent. This is more than GPT-6 Astra and Kimi K3.

Mistral’s stated goal is to lead the European sovereign AI market. The firm said ML4 was trained using 3,800 Nvidia Grace Blackwell GPUs hosted entirely within its European data centres, adding that its training data included every official EU language.

Founded in 2023, Mistral has emerged as a key player in European AI, particularly as France has worked to reduce its digital dependence on regions such as the US and expanding its domestic AI capabilities.

In September, Mistral announced it had raised €3 billion in a funding round led by Samsung Electronics, Scaleup Europe Fund and PSG Equity.

Its main competitor in Europe is Aleph Alpha, a German AI startup founded in 2019 which is undergoing a merger with the Canadian AI firm Cohere.

This week, Aleph Alpha announced the launch of Kolibri 1, a 78 billion parameter open-weight model optimised for German-language and mathematical tasks that competes with models such as Nvidia’s Nemotron, Google’s Gemma and OpenAI’s gpt-oss.

Aleph Alpha said that Kolibri 1 was trained “end-to-end from scratch by our teams in Germany, with full supply-chain integrity,” and that it is compliant with the EU AI Act.