Mistral AI unveils Large 4 AI model capable of processing different types of data

French company Mistral AI unveiled a new artificial intelligence model called Mistral Large 4 on Tuesday. Nicknamed “Le Chonk,” the model can process different types of data and has 1 trillion internal parameters. The company is presenting the model as an alternative to advanced systems developed in the United States and China. For now, it can be tested through a public testing service with safeguards in place.

Once safety testing is complete, the model’s weights, meaning the internal parameters it learned, are expected to be released within approximately 3 weeks. The final version is expected to be released by the end of the month. During this period, the company will work on safety with governments and trusted partners. Once the weights are released publicly, companies will be able to test, modify and adapt the model to their needs.

Mistral Large 4 was built from scratch and trained using the company’s own computing power. The model was trained with a total of only 4,000 NVIDIA GPUs. A GPU is a powerful computing chip that processes large amounts of data quickly. According to Mistral AI, the model required 2–3 times less computing power than its Chinese rivals. The resources used for closed models are considerably greater.

The model has been optimized for coding, cyber defense, finance, manufacturing, electrical engineering and chip design. Chip design is also an important field for the company’s main investors, ASML and Samsung. The final results of benchmark tests have not yet been announced. Mistral AI claims this will be the most powerful open-weight model developed outside China and will approach the performance level of some closed models.

The model was unveiled at a time of growing concern about access to advanced artificial intelligence systems. In June, the U.S. administration temporarily restricted the distribution of OpenAI and Anthropic models because of the risk of sophisticated cyberattacks. It was later revealed that the safety rules of some U.S. models had been bypassed and that the models had been used in attacks against companies and foreign government agencies. According to Mistral AI, an open model is less likely to be shut down without warning.

Businesses running open-weight models mainly pay for the computing power they use. Companies can host the model on their own servers, test its security and adapt it to specific tasks. Mistral AI charges for access to its cloud service. The company also earns revenue from the work of engineers who configure the models to meet customers’ needs.

The company received a $3.3 billion investment in September (approximately AZN 5.61 billion). During that investment round, the company was valued at approximately $24 billion (AZN 40.8 billion). Under another estimate, the figure was €21 billion (approximately AZN 40 billion). ASML led the Series C round, while Samsung led the Series D round. The company’s revenue is said to have increased approximately 20-fold over the past 1 year.