Mistral opens public preview of Large 4 model

Mistral announced a public preview of Large 4 and said the API was available through Mistral Studio at launch.[2] The company said it planned to release the model’s weights by the end of the month.[2] Documentation shown by Mistral is dated October 6, 2026.[1] Mistral calls Large 4 its largest and most capable model to date, and says it will become the foundation for a new generation of specialized and optimized Mistral models.[2]

Mistral describes Large 4 as an open-weight, general-purpose multimodal model using a Mixture-of-Experts architecture.[1] Its listed configuration includes 52 billion active parameters, 1.05 trillion total parameters and a 1.6-billion-parameter vision encoder.[1] The context window is listed as one million tokens.[3] Mistral says Large 4 will be available in multiple regions, including a European deployment it operates end-to-end under European law, and that the model will be able to run on private cloud or on-premises.[2]

Mistral’s documentation lists $1.36, $0.14 and $4.18 among Large 4’s prices.[4] Its changelog says launch pricing included a 50% discount for two weeks.[3] Mistral says the model is intended for cybersecurity, manufacturing, spatial reasoning, finance, multimodal use, retrieval-augmented generation and function calling.[5] The company says it worked with enterprises across finance, engineering, manufacturing, logistics, pharmaceuticals, science, shipping and the public sector during training.[2]

Mistral says Large 4 solved 93% of Cybench exercises and reports scores of 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4.[2] It reports 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase.[2] On Dense 200 visual grounding, Mistral reports 42%, compared with 41% for GPT-6-Astra.[2] Mistral also says the model resisted 93.3% of attacks on Lakera’s B3 AI Security Benchmark and had a higher average refusal rate on cyber prompts than the open-source models it compared.[2] These are company-reported results, so teams should validate performance on their own workloads.