Mistral Releases Open-Weight 1-Trillion-Parameter Model 'Large 4' on Hugging Face

Mistral has released Large 4 — nicknamed 'Le Chonk' — making it available for free download on Hugging Face. The one-trillion-parameter model was trained on Mistral's own 10-megawatt computing cluster in France. According to independent evaluator Artificial Analysis, Large 4 narrowly tops all other open-weight models on its Coding Agent Index, which measures how well AI agents complete full software engineering tasks. On Terminal-Bench 4.0, a separate coding-agent test, Large 4 ties with China's GLM-5.3, previously regarded as the world's best open model. Mistral says additional reinforcement learning — a training method in which the model receives feedback on task attempts and learns to favour what works — lifted the model's Terminal-Bench 4.0 score from 28 percent to 41 percent since a preview was released on 6 October. Large 4 also uses 47 percent fewer tokens per task than GLM-5.3. Proprietary models remain ahead overall: Anthropic's Mythos 5.1 scored 60.9 percent on the same benchmark. Mistral VP of science Pierre Stock called the release a turning point for European AI, noting the model was built on Mistral's own infrastructure — "a fraction of what all the competitors use." He argued that open-weight models offer businesses in finance, healthcare and manufacturing a safeguard against losing access to closed systems. That concern proved real in June, when a US export-control directive forced Anthropic to suspend access to two of its models for nearly three weeks. Once model weights are public, Stock said, access cannot be cut off in that way.