Mistral Large 4 Open

Changes from previous version

Replaces the retired Mistral Large 4 preview on the mistral-large-4 API after three more weeks of reinforcement learning. Terminal-Bench 4.0 on Mistral's chart rises from 28.3% to 41.4%. List price is $0.70 / $0.07 cached / $2.10, and the weights are public. High reasoning effort is the default, with medium and low also available.

Release Summary

Open-weight Mistral Large 4, released on 2026-10-29 and served as mistral-large-4 after Mistral retired the October 6 preview. The model card describes about 1 trillion total parameters, 49 billion active per token (52 billion including embeddings and output layers), text and image in, and text out. Context is 1M tokens. List price is $0.70 per million input tokens, $0.07 cached, and $2.10 output. A measured price per task at high effort, the API default, is $2.06. Reasoning effort is low, medium, or high. On Mistral's charts, Terminal-Bench 4.0 is 41.4%, up from 28.3% for the preview, FrontierCode 2.0 is 19.7%, and the Coding Agent Index is 48% at high effort (8.5 million average tokens per task) and 35% at medium. The same Coding Agent Index chart lists GLM-5.3 at 49% and 15.9 million tokens. Cyber Index is 50.9%, BrowseComp is 84.5%, and FinWorkBench (Finch) is 73.3%. Mistral also reports 30% fewer Terminal-Bench 4.0 output tokens than the preview. Weights are on Hugging Face as mistralai/Mistral-Large-4-1T-A52B under the Mistral Ecosystem Open-Weight license. A Model-as-a-Service business needs a separate commercial license. DSpark is a 3-layer draft head on these same weights. Also compare Kimi K3 and DeepSeek V4.1-Flash-1017.

Timeline

Other Mistral Large versions