Mistral open-sourced the trillion-parameter Large 4 model weights, reporting that its programming test scores rose by 13 percentage points while per-question costs dropped by 63%.

Beating AI News Flash: French AI company Mistral has officially open-sourced the weights of its flagship model Large 4 (codenamed Le Chonk). The model has a total of 1.05 trillion parameters, with approximately 49 billion activated per inference. After opening API previews on October 6, Mistral conducted three additional weeks of reinforcement learning training, and the version released now is an improved iteration. Mistral reported its Terminal-Bench 4.0 programming score rose from 28% to 41%; FrontierCode 2.0 scored 19.7%, ranking first among open-source models. In Artificial Analysis' programming agent index, Large 4 scored 48%, close to GLM-5.3's 49%, while consuming 47% fewer tokens per average question. However, Artificial Analysis only recorded a 25% score for Terminal-Bench 4.0 under its own test configuration, failing to replicate the official 41% result. In other tests, financial table task Finch's score increased from 67.4% to 73.3%; the agent's success rate in resisting prompt injection attacks rose from 93.3% to 97.1%. Its cybersecurity index stands at 51%, still lower than DeepSeek V4.1 Flash's 57%, with more modest overall capability improvements. Artificial Analysis has just upgraded its intelligence index to v5. On the same metric, Large 4's score rose from 25 in the preview version to 26, below GLM-5.3 and DeepSeek V4.1 Flash's 32 points. But inference efficiency improved significantly: output tokens per question dropped by 39%, and hallucination rate fell from 41% to 28%. API prices were also nearly halved. Based on October 29 evaluation pricing and usage estimates, per-question cost fell from $5.54 to $2.06, a 63% reduction. The model weights have been uploaded to Hugging Face, offering FP8 and NVFP4 versions, supporting self-deployment, fine-tuning, and general commercial use. Separate authorization from Mistral is required to operate a commercial model hosting service.