Mistral releases open weights for Large 4 AI model

Mistral AI released the open weights for its Large 4 model on October 29, three weeks after previewing it, saying additional training improved its coding performance. The company describes the model, nicknamed Le Chonk, as a trillion-parameter system with about 52 billion parameters active for each task. It accepts text and images and generates text and code. Mistral said the model’s TerminalBench 4.0 score rose from about 28 at preview to 41 after three further weeks of reinforcement learning. The benchmark tests coding agents working in a terminal. Mistral also reported that Large 4 used 47% fewer tokens per task than GLM5.3 on the Artificial Analysis Coding Agent Index, where it scored 48%. Those results do not establish broad leadership across model evaluations. Artificial Analysis gave Large 4 an Intelligence Index score of 26. The Decoder reported that this placed it fifth among open models, with Chinese models scoring higher. On the coding index, The Verge reported that Large 4 edged GLM5.3 on TerminalBench by one point, while its composite score was one point lower. Results therefore vary by benchmark and measure. Mistral offers low, medium and high reasoning modes,...