Mistral Small 3 brings open source AI into the mass small, faster and cheaper

Mistral they haveThe rapidly rising European startup for artificial intelligence today presented a new voice model that claims that it corresponds to the performance of models three times its size and at the same time reduces the computing costs – a development that could convert the economy of the extended AI deployments.

The new model, named Mistral Small 3Has 24 billion parameters and reaches 81% accuracy for standard benchmarks, while 150 tokens are processed per second. The company publishes it under the revealing Apache 2.0 licenseSo that companies can change and provide it freely.

“We believe that it is the best model among all models of less than 70 billion parameters,” said Guillaume Lampe, Chief Science Officer from Mistral, in an exclusive interview with Venturebeat. “We estimate that it basically corresponds to the Lama 3.3 70b of the meta that was published a few months ago. This is a model that is three times larger.”

The announcement comes in the middle Intensive examination of AI development costs according to the Chinese startup Deepseek that it has developed a competitive model Only $ 5.6 million – claims that have wiped off Almost 600 billion US dollars From Nvidia’s market value this week when investors questioned the massive investments of US -Tech giants.

Like a French startup a AI

Mistral’s approach focuses more on efficiency than on scale. The company reached its performance gains mainly through improved training techniques instead of throwing more computing power on the problem.

“What has changed is basically the training optimization techniques,” Lamp told Venturebeat. “The way we train the model was a little different, another way to optimize it, change the weights during free learning.”

According to the lamp, the model was trained on 8 trillion tokens compared to 15 trillion for comparable models. This efficiency could make advanced AI functions more accessible to companies that are concerned about the computing costs.

Above all, Mistral Small 3 Was developed without strengthening learning or synthetic training data, techniques that are usually used by competitors. According to the lamp, this “raw” approach helps to avoid embedding unwanted distortions, which could be difficult to recognize later.

Privacy and company: Why do smaller AI models do for mission-critical tasks

The model is particularly aimed at companies that are used for reasons of privacy and reliable reasons, including financial services, healthcare and manufacturing companies. According to the company, it can run on a single GPU and carry out 80-90% of the typical corporate use cases.

“Many of our customers want a local solution because they take care of privacy and reliability,” said Lample. “You don’t want critical services that rely on systems that you do not fully control.”

The Europe’s KI Champion is preparing the stage for Open -Source Dominance as an IPO

The publication is Mistral, worth 6 billion US dollarspositions itself as a European champion in the global AI race. The company recently invested from Microsoft and is preparing for one eventual IPOAccording to CEO Arthur Mensch.

Industry observers say that Mistral’s focus could be on smaller, more efficient models as foresight if the AI ​​industry matures. The approach is in contrast to do so Openai And Anthropic This has focused on developing increasingly large and expensive models.

“We will probably see the same thing we saw in 2024, but maybe even more than what basically many open source models with very permissible licenses,” said Lampe. “We believe that it is very likely that this conditional model has become a kind of goods.”

Since the competition is intensified and the efficiency gains arise, Mistral’s strategy could optimize smaller models, contribute to democratizing access to advanced AI skills – potentially accelerates acceptance in the industries and at the same time the costs for computer infrastructures.

The company says it will publish additional models with Enhanced Argumentation skills In the coming weeks, an interesting test will be created for whether its efficiency -oriented approach can continue to meet the skills of much larger systems.