Mistral Opens Weights on 1T-Param Large 4 Model
Original: We're releasing the weights of Mistral Large 4
Why This Matters
Open 1T-parameter models with frontier coding scores raise the bar for enterprise self-hosting.
Mistral released open weights for Mistral Large 4 (Le Chonk), a 1-trillion-parameter multimodal model, topping FrontierCode 2.0 at 19.7% and matching rival GLM-5.3 using 47% fewer tokens.
Mistral put its flagship model's weights into the open three weeks after the initial public preview of Mistral Large 4, nicknamed Le Chonk. The 1-trillion-parameter multimodal model leads the open-weight field on Cognition's private FrontierCode 2.0 benchmark with a score of 19.7%, ahead of DeepSeek V4.1 Flash (19.3%) and GLM-5.3 (17.8%). On the Artificial Analysis Coding Agent Index, it scores 48%—effectively tied with GLM-5.3—but uses 47% fewer tokens per task, a meaningful cost advantage for developers running coding agents at scale. Mistral also claims top marks on Terminal-Bench and strong results in defensive cybersecurity, areas it says continued to improve through three additional weeks of reinforcement learning after the preview launch, with 'no signs of saturation.' The model offers three reasoning modes to let users trade off cost against performance. Trained entirely on European infrastructure with full lineage tracking, Mistral is positioning the release as a building block for sovereign AI—deployable on-premises or in self-hosted environments. Weights are available now.