Mistral releases improved Large 4 weights with reasoning-efficiency gains

Mistral shipped Large 4 in preview on October 6, and the weights arrived on October 29.[1][2] Mistral describes the released model as improved, with gains in reasoning efficiency.[3] Large 4 is a natively multimodal mixture-of-experts model with about 1 trillion total parameters.[4] Mistral’s model card gives 49 billion active parameters per token, or 52 billion when embeddings and output layers are included.[4] The release makes the weights available, but the evidence does not specify additional deployment terms beyond the licensing note for Model-as-a-Service use.[4]

Large 4 accepts text and image input and generates text, and Mistral says it supports a 1-million-token context window.[4] The model has low, medium and high reasoning settings, with high as the default.[4] Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European datacenters and is natively fluent in more than 160 languages, including every official EU language.[4] Its listed API rates are $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.[5] Operating a Model-as-a-Service business requires a separate commercial license with Mistral.[4]

Mistral says Large 4 scores 48% on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens per task on average.[4] On Terminal-Bench 4.0, Mistral reports 41%, up from 28% in the public preview.[3] Mistral also reports a 19.7% score on Cognition’s FrontierCode 2.0 benchmark.[3] For cybersecurity, Mistral reports a 51% score on the Artificial Analysis Cyber Index and a 96.6% average refusal rate on malicious cyber requests, which it says is the highest it measured among open models.[3] These are reported benchmark results, useful as inputs to evaluation rather than substitutes for testing against a team’s own tasks.

Mistral says enterprise design partners, including Orange Cyberdefense, have preview access to a specialized variant for advanced cybersecurity workflows.[3] Testing that variant is restricted to owned or explicitly authorized systems in controlled environments.[3] Mistral says the model matched vulnerability-detection recall rates against other frontier cyber models while offering a compelling cost curve.[3] It also says it sees potential for uplift to conduct real attacks and is modifying its guardrails.[3]