For the latest updates and exclusive contents of the industry’s best AI application, join the daily and weekly newsletter. Learn more
Mistral AI, a rapidly rising European AI start -up, has released a new language model today, which matches the performance of three times the performance and dramatically reduces computing costs.
The new model, Mistral Small 3, has 24 billion parameters, with 150 tokens per second to achieve 81% accuracy on the standard benchmark. The company can disclose this in accordance with the allowed Apache 2.0 license so that the company can freely modify and distribute.
Guillaume Lample, the chief scientific officer of Mistral, said in an exclusive interview with Venturebeat, “We think it’s the best model of all models of less than 7 billion parameters. “We are basically estimated to be equivalent to the meta’s LLAMA 3.3 70b, which was released a few months ago. This is three times larger.”
The announcement trained a competitive model of only $ 55 million of the intense investigation of AI development costs, according to China’s start -up DeepSeek. Giants.
A method in which French startups build an AI model that competes with large technology in some of the size
The approach to Mistral focuses on efficiency, not scale. The company did not throw more computing power on the problem and gained performance benefits mainly through improved educational technology.
Lample said to VentureBeat, “Basically, training optimization technology. “The way we train the model is slightly different to optimize the model. We have modified the weight during free learning.”
According to Lample, this model has been trained in 8 trillion tokens with 15 trillion tokens for similar models. This efficiency makes it easier to access advanced AI features in business related to computing costs.
In particular, Mistral Small 3 was developed without reinforcement learning or synthetic education data, and the technology commonly used by competitors. Lample said that this “primitive” approach helps prevent them from inclusing unwanted prejudice later.
Privacy and Enterprise: Reasons why businesses are watching small AI models for mission critical work
This model is especially for companies that require on-premises deployment for personal information and reliability, including financial services, medical and manufacturers. According to the company, it is executed in a single GPU and can handle 80-90%of general business use cases.
Lample said, “Many customers are interested in the protection and reliability of personal information. “They do not want an important service that relies on a system that does not completely control.”
European AI champions set the stage of open source dominance by IPO straight.
This release is a $ 6 billion Mistral, a European champion in the global AI race. According to CEO of Arthur Mensch, the company has recently invested from Microsoft and is preparing for the final IPO.
Observations in the industry say that the focus of Mistral on small and efficient models can be proved to be prerequisites as the AI industry matures. This approach contrasts with companies such as Openai and Anthropic, which focuses on developing more and more expensive models.
Lample said, “We will see the same thing as we saw in 2024. But perhaps you can see more than this. Basically, many open source models with licenses allowed. “We think this conditional model is very likely to be a kind of product.”
As competition intensifies and increases efficiency, Mistral’s strategy, which optimizes small models, can help democratize access to advanced AI functions.
The company will announce additional models with strengthened reasoning functions in the next few weeks, and will establish interesting tests on whether the efficiency -centered approach can continue to match the function of the system.
