Mistral open-sources trillion-parameter multimodal model Mistral Large 4 (Le Chonk)

On October 29, 2026, Mistral released and open-sourced the weights of its flagship model Mistral Large 4 (nicknamed Le Chonk) via Hugging Face and its official blog. The model uses a MoE architecture with roughly 1 trillion (1T) total parameters, a 1M-token context window, and supports text and image input. It is licensed under the Mistral Ecosystem Open-Weight License and is available on Hugging Face. Mistral AI's official blog calls it a 1-trillion-parameter natively multimodal open-weight model, focused on coding and cybersecurity, improved through three weeks of reinforcement learning after a public preview. The Hugging Face Blog lists the activated parameters as 49B, while a later report from The Decoder states 52B activated per step — the two figures are inconsistent. The Decoder also noted that the model leads open-source models on coding tasks but still trails Chinese models. Apart from the discrepancy in activated-parameter figures, the reports consistently describe the model's basic positioning (1T parameters, multimodal, open weights).

AI overview of reporting · Updated 1 h前

Not enough continuous observations to show a trend yet.