Mistral opens public preview of Large 4

Mistral says the public preview is available through an API on Mistral Studio, and that it will release the model weights by the end of the month.[1] Its documentation describes Large 4 as an open-weight, general-purpose multimodal model built with a granular Mixture-of-Experts architecture.[3] The documentation lists 52 billion active parameters and 1.05 trillion total parameters, alongside a 1.6-billion-parameter vision encoder.[3] Teams can assess the preview against their own workloads now, while treating access to the weights as a separate upcoming step.

Mistral lists API prices of $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.[3] Its documentation gives the model a one-million-token context length and says it supports retrieval-augmented generation and function calling.[3][4] Mistral says customers can run Large 4 on private cloud or on-premises infrastructure.[1] These details give technical and procurement teams concrete variables to test: expected input and output volume, context needs, and which deployment arrangement fits their requirements.

Mistral says Large 4 is designed for tasks including cybersecurity, manufacturing, spatial reasoning, finance and multimodal use.[4] The company reports a 73.3% score on the Finch finance benchmark, second among open models to Kimi K3, and an 84.5% score on BrowseComp, up from 75% for its starting supervised-fine-tuning checkpoint.[2] Mistral also reports 42% on the Dense 200 visual-grounding benchmark, compared with 41% for GPT-6 Astra.[1] These are Mistral-reported benchmark results; teams should test whether they reflect their own use cases.

Mistral says it trained Large 4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers.[1] It says a significant share of the training data was multilingual, spanning more than 160 languages, including every official EU language.[1] On safety, Mistral reports a 96.6% average refusal rate on malicious cyber requests, which it says was the highest it measured among open models.[2] The company also reports a 51% score on the Artificial Analysis Cyber Index.[2] For evaluation, teams can treat these figures as reasons to include relevant multilingual and security tests, not as substitutes for them.