Mistral's Pierre Stock on Le Chonk: trillion-parameter open model trained on 3,800 GPUs, RL still running, Large 5 in pretraining

On The MAD Podcast with Matt Turck (22 October), Mistral VP of Science and first employee Pierre Stock discussed Mistral Large 4, nicknamed 'Le Chonk', a 1-trillion-parameter open-weight mixture-of-experts model trained in two months on 3,800 GPUs in Mistral's own European data centres. Stock said the reinforcement-learning run is still going, that Mistral Large 5 is already in pretraining and that 10-trillion-parameter models are coming. He argued the US move to restrict access to closed models for non-US users showed AI access 'could be cut off' like electricity, strengthening the case for open weights. Mistral says weights are due by end-October.