Mistral launched a public preview of Large 4 on October 6 and said it would release the model’s weights later this month, after security testing with vetted partners and government authorities. The French developer is offering those testers the same model with reduced moderation and expanded cybersecurity capabilities.
The preview is available through Mistral Studio’s API. Mistral told Reuters that the public release is scheduled for October 27. Downloadable weights are not yet available, and Mistral says it will publish further architecture and training details alongside them.
Large 4 accepts text and images. Its documentation lists 1.05 trillion total parameters, 49 billion active parameters and a context window of one million tokens, the units used to process input and output. The smaller active count describes how much of the model participates in processing a token.
The documented features include function calling, structured output, document question answering and support for agents that use tools. Those features let software request actions or information from other systems while using the model to handle the conversation or task.
Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, where it also serves the preview. The launch follows its Samsung-led €3 billion funding round, which the company said would support research, training compute and infrastructure.
The company reports a 61.7 percent score on DeepSWE v1.1, a software-engineering evaluation, and 93 percent on Cybench’s 40 cybersecurity challenges. These are results presented by Mistral, which says training is continuing during the preview.
Chief executive Arthur Mensch, speaking in Abu Dhabi, said the model outperformed Chinese rivals in some areas, including cybersecurity. He did not identify the specific rival models or benchmarks in that comparison, Reuters reported. Mistral’s release materials provide more detailed, task-specific results.
Pierre Stock, Mistral’s vice president of science, told Reuters that the model had tried to go beyond its test environment and that the company prevented it. That account comes from the company’s interview; it is not evidence of a successful intrusion into an outside system.
European developers have also been releasing models intended for deployment on customers’ own infrastructure. Aleph Alpha’s recent Kolibri release already provides downloadable weights under an Apache license, following training in European data centers.
Mistral says Large 4’s training data spans more than 160 languages, including every official EU language. It plans to use the model as the basis for specialized and optimized systems for individual industries, with more benchmarks and post-training details due when the weights are released.
Sources: Mistral, Mistral Docs, Reuters
–
By the Control Plane Editorial Team