Mistral AI has launched its Mistral Large 4 model in Research Public Preview. The model features 1 trillion parameters with 49 billion active parameters and supports a 512k token context window. It is currently available via the company's API, with open weights expected to be released by the end of October.
On the Artificial Analysis Intelligence Index, the model achieved a score of 38. While Mistral identifies it as the most intelligent model outside the US and China, it trails several Chinese open-weight models, including Kimi K3 and GLM-5.3-Flash. The model demonstrates strong performance in cybersecurity benchmarks, specifically scoring 82% on CyberGym-E2E-AA.
The standard pricing for the model is $1.36 per 1 million input tokens and $4.18 per 1 million output tokens. For the first two weeks of the preview, Mistral is offering a 50% discount on these rates. The cost per task is currently higher than that of comparable open-weight models like GLM-5.3-Flash and DeepSeek V4.1 Flash.
Mistral trained the model from scratch using 3,800 NVIDIA Grace Blackwell GPUs in its European data centers. The training process took approximately two months. This infrastructure is part of the company's broader expansion, which includes plans to utilize up to 18,000 Grace Blackwell chips to support future specialized model development.
