Mistral Large 4

Changes from previous version

The next cataloged Mistral Large after Mistral Large 2. Adds native image input, a 1M context window, and the mistral-large-4 preview API. On Mistral's charts it leads open models outside China on CyberGym-E2E at 82% and edges GPT Astra 6 on Dense200 (bbox), 42.0% to 41.5%.

Release Summary

Public preview of Mistral's largest model, served as mistral-large-4. The announcement describes a natively multimodal mixture-of-experts with about 1 trillion parameters and 49 billion active. The model card lists 1.05 trillion total parameters, 52 billion active, and a 1.6 billion vision encoder. Text and image in, text out, with a 1M-token context window. List price is $1.36 per million input tokens, $0.14 cached, and $4.18 output. The first two weeks from 2026-10-06 are 50% off ($0.68 / $0.07 / $2.09). A measured price per task is $1.13 at list rates, and $0.57 during that discount. Mistral reports 82% on CyberGym-E2E, 93% on Cybench, 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4.0, and 42.0% on Dense200 (bbox) versus 41.5% for GPT Astra 6. Weights are planned for the end of October 2026. Also compare Claude Opus 5, Kimi K3, and GLM-5.3.

Timeline

Other Mistral Large versions