On October 6th, according to news from the IT community, French company Mistral AI officially launched a public preview version of the Mistral Large 4 model (Public Preview), commonly referred to as ML4, with the code names le and Chonk.
Starting from today, users can experience the preview version of API on Mistral Studio. The model weights will be officially available for download at the end of this month.
Currently, the official team is collaborating with top cybersecurity organizations, trusted partners, and government regulatory authorities to conduct red-team testing on the model in a real business environment ( Red-Teaming ). The relevant partners will be granted special access permissions for testing this model; this test version has appropriately relaxed the content security review policies and enhanced the capabilities in cybersecurity attack and defense techniques.
ML4 is a native multimodal model with a total of 1 trillion parameters and 49 billion activated parameters. Mistral claims that its actual test performance is not only comparable to the world's top open-source models but also significantly ahead of all open-weight models developed in Europe and America.
In critical enterprise-level workloads such as network security, finance, and law, this model has achieved SOTA among all open-source models. In certain specific areas, such as visual localization, its performance goes even further, surpassing even the current state-of-the-art proprietary models.
According to the introduction, ML4 is trained from scratch using 3,800 NVIDIA Grace Blackwell GPU within the self-built data center of Mistral located in Europe. Its European deployment plan is entirely independently operated end-to-end by Mistral, strictly complying with EU laws and regulations, and is not subject to any other third-party digital service providers.
The training data of ML4 possesses significant multilingual characteristics, with a corpus covering over 160 natural languages, encompassing all official languages of the European Union.
In terms of price, ML4 costs $1.36 per million tokens inputs (which is approximately 9.1 RMB at the exchange rate mentioned in the text), and it yields $4.18 per million tokens outputs (which is approximately 28.1 RMB at the same exchange rate).
