Table of Contents
Become a member of our daily and weekly newsletters for the latest updates and exclusive content about leading AI coverage. Leather
Mistral AIThe rapidly rising European startup of artificial intelligence has today unveiled a new language model that it claims to be the performance of models three times its size, while the calculation costs are drastically reduced – a development that could reform the economy of advanced AI implementation.
The new model, called Mistral Small 3Has 24 billion parameters and reaches 81% accuracy on standard benchmarks during the processing of 150 tokens per second. The company is publishing it under the toll mission Apache 2.0 licenseCompanies allow it freely to change and implement it.
“We believe that this is the best model between all models of fewer than 70 billion parameters,” said Guillaume Lample, Chief Science Officer of Mistral, in an exclusive interview with Venturebeat. “We estimate that it is in fact the same as the Lama 3.3 70b of the meta that was released a few months ago, which is three times larger.”
The announcement comes in the midst of intensity of AI development costs after claims from the Chinese startup Deepseek that it has trained a competitive model for Only $ 5.6 million – claims that wipe Almost $ 600 billion Of the market value of Nvidia this week when investors questioned the enormous investments made by American technical giants.
How a French startup has built an AI model that Big Tech competitions a fraction of the size
Mistral’s approach focuses on efficiency instead of scale. The company mainly achieved its performance gain through improved training techniques instead of throwing more computing power to the problem.
“What changed is actually the training optimization techniques,” Lample told Venturebeat. “The way we train the model was a bit different, another way to optimize it.”
The model was trained on 8 trillion tokens, compared to 15 trillion for similar models, according to Lample. This efficiency can make advanced AI options more accessible to companies that are concerned about the calculation costs.
Remarkable, Mistral Small 3 Was developed without reinforcement learning or synthetic training data, techniques that are often used by competitors. Lample said that this “rough” approach helps to bed in unwanted prejudices that can be difficult later.
Privacy and Enterprise: Why companies look for smaller AI models for mission-critical tasks
The model is particularly aimed at companies that require the use of on-premises for reasons for privacy and reliability, including financial services, health care and production companies. According to the company, it can be carried out on one GPU and process 80-90% of the typical business user situations.
“Many of our customers want an on-premises solution because they care about privacy and reliability,” Lample said. “They don’t want critical services dependent on systems that they do not fully control.”
The AI champion of Europe is the scene for open source dominance when IPO looms up
The release comes as Mistral, appreciated at $ 6 billionPositions itself as Europe’s champion in the global AI race. The company recently made investments by Microsoft and is preparing for one Possible IPOAccording to CEO Arthur Mensch.
Industry observers say that Mistral’s Focus on smaller, more efficient models could be looking forward to as the AI industry becomes mature. The approach is in contrast with companies such as Openi And Anthropic who have focused on developing ever larger and expensive models.
“We will probably see the same thing that we have seen in 2024, but perhaps even more than this, which is in fact a lot of open-source models with very permitted licenses,” Lample predicted. “We believe that it is very likely that this conditional model has become a kind of raw material.”
As the competition increases and the efficiency transactions arise, Mistral’s strategy to optimize smaller models can help democratize access to advanced AI options – possibly speeding up acceptance in industry and at the same time lowering the costs of computer infrastructure.
The company says that it will release additional models with improved reasoning options in the coming weeks, so that an interesting test is set up or the efficiency -oriented approach can continue to match the possibilities of much larger systems.