European AI flag bearer Mistral's new open weights model is 'Le Chonk'

Mistral, Europe's AI leader, has unveiled its latest open weights model, dubbed "Le Chonk," a 1 trillion-parameter large language model (LLM). The French company has christened the model Mistral Large 4, which is bound to generate much interest. Standing in the frontier territory of AI, Le Chonk is multimodal, reasoning, and utilizes a mixture-of-experts (MoE) architecture.

While the model's hardware requirements aren't specified, it's anticipated to run efficiently on 8-way GPU boxes like Nvidia's HGX B300 or AMD's MI355X.

For training, Mistral used around 3,800 Grace Blackwell GPUs - approximately 52 NVL72 racks - in their European datacenters. This model, trained on a corpus spanning 160 languages, combines supervised pre-training and reinforcement learning. Le Chonk is currently in preview, with its weights expected to become available on Hugging Face and other model repositories within a month.

Early independent testing shows Le Chonk trailing behind top AI models from OpenAI and Anthropic. However, it surpasses the most capable open weights model, Thinking Machines Lab's Inkling, from the United States. In terms of non-Chinese LLMs available for download, Mistral holds its own quite well. Their self-benchmarks, though, paint a rosier picture across various domains, including coding, agentic, finance, legal, and cybersecurity tasks.

Mistral is particularly proud of Le Chonk's strong performance in cybersecurity workloads, which they use to counteract proprietary American AI models with such restrictions. Their €3 billion Series D funding round last month aims to propel Mistral into developing more advanced models, with the expectation of performance improvements in future iterations of Le Chonk. The model is currently accessible via Mistral's API.