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

European AI leader Mistral has unveiled a monumental new open weights model named Le Chonk, after the company jokingly dubbed it. This 1 trillion-parameter large language model (LLM), officially titled Mistral Large 4, marks Mistral's most extensive open weights release to date. As a frontier model, Le Chonk boasts multimodal reasoning and a mixture-of-experts (MoE) architecture with approximately 49 billion active parameters, enabling efficient serving despite its size.

While Mistral hasn't disclosed hardware recommendations, the model is compatible with 8-way GPU configurations like Nvidia's HGX B300 and AMD's MI355X.

Mistral trained Le Chonk using over 3,800 Grace Blackwell GPUs, equivalent to around 52 NVL72 racks, within their European data centers on a multilingual corpus exceeding 160 languages through supervised pre-training and reinforcement learning. This sovereign AI effort aims to deliver frontier-class models within Europe. Currently in preview, the model's weights will be available on Hugging Face and other model repositories within a month.

Initial testing by Artificial Analysis places Le Chonk below OpenAI's GPT6 Luna and Anthropic's flagship models, but surpasses the most capable American open weights model, Thinking Machines Lab's Inkling.

Performance-wise, Le Chonk holds its own against leading Chinese LLMs from Alibaba, Moonshot, Z.AI, and DeepSeek in various benchmarks, including coding, finance, and legal tasks. Notably, Mistral touts strong cybersecurity performance in Le Chonk, countering proprietary American models that often restrict such adversarial tasks. With Mistral's €3 billion Series D funding round behind it, the company anticipates ongoing performance enhancements, potentially leading to an improved Mistral Large 4.1 in the near future.