Mistral Large 4
Points and comments are a snapshot, not live.
Mistral releases Large 4, a 1-trillion-parameter open-weight multimodal model.
Mistral Large 4 (ML4, aka "le Chonk") is a 1-trillion-parameter multimodal model with 52 billion active parameters, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in European datacenters. It achieves top-five scores on the Artificial Analysis Cyber Index and 82% on a real-vulnerability reproduction test. The model scores 61.7% on DeepSWE v1.1 and ranks second in a blind human coding evaluation. A public preview API is available now; weights will be released by the end of October. The model uses reinforcement learning at scale, generating roughly 33 billion tokens per day during training.
What commenters are saying
Commenters note that Mistral Large 4 is weaker and more expensive than DeepSeek models on benchmarks, with the main differentiators being European data sovereignty and open weights. Many point out that Mistral itself hosts Chinese models like GLM-5.3 in Europe, reducing the practical difference. Several skeptics see patriotism or regulatory compliance as the only reason to choose this model. Some users report the model spits out Chinese characters in German conversation, raising speculation about training data composition or distillation. Two camps: those who value European control and those who see it as performance without benefit.