Mistral Small 3 brings open-source AI to the masses — smaller, faster and cheaper

by

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More

Mistral AI, the rapidly ascending European artificial intelligence startup, unveiled a new language model today that it claims matches the performance of models three times its size while dramatically reducing computing costs — a development that could reshape the economics of advanced AI deployment.

The new model, called Mistral Small 3, has 24 billion parameters and achieves 81% accuracy on standard benchmarks while processing 150 tokens per second. The company is releasing it under the permissive Apache 2.0 license, allowing businesses to freely modify and deploy it.

“We believe it is the best model among all models of less than 70 billion parameters,” said Guillaume Lample, Mistral’s chief science officer, in an exclusive interview with VentureBeat. “We estimate that it’s basically on par with the Meta’s Llama 3.3 70B that was released a couple months ago, which is a model three times larger.”

The announcement comes amid intense scrutiny of AI development costs following claims by Chinese startup DeepSeek that it trained a competitive model for just $5.6 million — assertions that wiped nearly $600 billion from Nvidia’s market value this week as investors questioned the massive investments being made by U.S. tech giants.

Mistral Small 3 achieves similar performance to larger models while operating with significantly lower latency, according to company benchmarks. The model processes text nearly 30% faster than GPT-4o Mini while matching or exceeding its accuracy scores. (Credit: Mistral)

How a French startup built an AI model that rivals Big Tech at a fraction of the size

Mistral’s approach focuses on efficiency rather than scale. The company achieved its performance gains primarily through improved training techniques rather than throwing more computing power at the problem.

“What changed is basically the training optimization techniques,” Lample told VentureBeat. “The way we train the model was a bit different, a different way to optimize it, modify the weights during free learning.”

The model was trained on 8 trillion tokens, compared to 15 trillion for comparable models, according to Lample. This efficiency could make advanced AI capabilities more accessible to businesses concerned about computing costs.

Notably, Mistral Small 3 was developed without reinforcement learning or synthetic training data, techniques commonly used by competitors. Lample said this “raw” approach helps avoid embedding unwanted biases that could be difficult to detect later.

In tests across human evaluation and mathematical instruction tasks, Mistral Small 3 (orange) performs competitively against larger models from Meta, Google and OpenAI, despite having fewer parameters. (Credit: Mistral)

Privacy and enterprise: Why businesses are eyeing smaller AI models for mission-critical tasks

The model is particularly targeted at enterprises requiring on-premises deployment for privacy and reliability reasons, including financial services, healthcare and manufacturing companies. It can run on a single GPU and handle 80-90% of typical business use cases, according to the company.

“Many of our customers want an on-premises solution because they care about privacy and reliability,” Lample said. “They don’t want critical services relying on systems they don’t fully control.”

Human evaluators rated Mistral Small 3’s outputs against those of competing models. In generalist tasks, evaluators preferred Mistral’s responses over Gemma-2 27B and Qwen-2.5 32B by significant margins. (Credit: Mistral)

Europe’s AI champion sets the stage for open source dominance as IPO looms

The release comes as Mistral, valued at $6 billion, positions itself as Europe’s champion in the global AI race. The company recently took investment from Microsoft and is preparing for an eventual IPO, according to CEO Arthur Mensch.

Industry observers say Mistral’s focus on smaller, more efficient models could prove prescient as the AI industry matures. The approach contrasts with companies like OpenAI and Anthropic that have focused on developing increasingly large and expensive models.

“We are probably going to see the same thing that we saw in 2024 but maybe even more than this, which is basically a lot of open-source models with very permissible licenses,” Lample predicted. “We believe that it’s very likely that this conditional model is become kind of a commodity.”

As competition intensifies and efficiency gains emerge, Mistral’s strategy of optimizing smaller models could help democratize access to advanced AI capabilities — potentially accelerating adoption across industries while reducing computing infrastructure costs.

The company says it will release additional models with enhanced reasoning capabilities in the coming weeks, setting up an interesting test of whether its efficiency-focused approach can continue matching the capabilities of much larger systems.



Leave a Comment