Mistral previews Large 4 'Le Chonk', a ~1T-parameter open-weight model trained in Europe

Mistral previewed Mistral Large 4, nicknamed "Le Chonk." It is a mixture-of-experts model with roughly 1 trillion total parameters and 49 billion active per token, trained on 3,800 NVIDIA Grace Blackwell GPUs in data centers that Mistral itself operates in Europe. Mistral claims it is among the strongest open-weight systems available, with improvements in coding, agentic workflows and multimodal understanding. The preview follows the October 20 release of Voxtral 3 speech models and an Agent SDK.

The architecture matters for deployment economics. With 49B active parameters, inference cost per token is closer to a mid-size dense model, even though total capacity rivals frontier systems. Training on Mistral-operated European infrastructure is a sovereignty pitch to EU governments and regulated industries wary of US and Chinese clouds.

Competitively, Mistral faces pressure from both sides. Chinese open-weight models are cheaper. Ecosia just reported cutting AI costs ~50% by dropping Mistral for Qwen, GLM and others. Meanwhile Meta's Muse Spark Open and DeepSeek's V4.1 compete for open-weight mindshare. Large 4 is Mistral's argument that it can still play at the top. A separate same-day NVIDIA partnership on industrial engineering, which CEO Arthur Mensch called "a key pillar of our research," reinforces its enterprise niche.

Caveats: this is a preview, and performance claims come from Mistral without independent benchmarks. Running a 1T-parameter model requires serious hardware even with sparse activation. Watch the license terms, the release date, and pricing relative to Qwen.