The fast open AI model Mistral Small 3 is presented-it can be launched on MacBook or GeForce RTX 4090

The French company Mistral AI, founded by immigrants from Google Deepmind and Meta✴, introduced the compact model of artificial intelligence Mistral Small 3 – it spreads with open code and offers high performance.

Image Source: Michael Dziedzic / Unsplash.com

The Mistral Small 3 has 24 billion parameters and is positioned as a direct competitor to larger models, including META✴ LLAMA 3.3 70B and Alibaba QWEN 32B, as well as the free replacement of closed systems, including OpenAi GPT-4O Mini. The developer assures that the new model gives the results at the same level with Llama 3.3 70b Instructor, but it works three times faster on the same equipment, and it is designed for 80 % of the tasks of generative AI.

Hereinafter, the source of the image: mistral.ai

The system was created with an eye on local deployment – its architecture has significantly smaller layers in comparison with competing models, which reduces the time for direct passage. The accuracy of the Mistral Small 3 in the MMLU test is 81 % with a delay in 150 tokens per second, which, according to the developer, makes it the most effective in its category. It can also serve as the basis for creating more complex reasoning models, such as Deepseek R1.

The new neural network was tested blindly among third -party contractors – the procedure included more than a thousand tasks to write code and answers to general questions. Mistral Small 3 issued competitive results in comparison with open models of three times more, as well as a closed GPT-4O Mini in trials related to writing code, solving mathematical problems, testing general knowledge and performing instructions.

In practice, the model will be useful in cases where quick and accurate answers are critically important-these can be virtual assistants, AI-agents and automation of work processes in real time. The developer has provided for users the ability to carry out a fine setting up Mistral Small 3 for its specialization in certain areas – these can be legal advice, medical diagnostics and technical support.

In the field of medical services, Mistral Small 3 is able to identify fraud attempts; in medicine – to direct patients to the right specialists; in robotics, auto industry and production – carry out control and control functions; Virtual customer service scenarios, mood analysis and reviews are provided. It can be launched on a system with one Nvidia GeForce RTX 4090 video card or on the actual Apple MacBook with 32 GB of RAM.

The Mistral Small 3 is already available and will appear on all the largest specialized platforms in the near future – it spreads through the free and open Apache 2.0 license.