The French artificial intelligence corporation Mistral AI announced on the 30th of last month (local time) the release of the optimized 24 billion parameter artificial intelligence (AI) model "Mistral Small 3." The model will be distributed under the Apache 2.0 license.
According to the company, Mistral Small 3 can compete with large models such as Meta's Llama 3.3 70B and Alibaba's Qwen 32B. Mistral AI noted that it is evaluated as a powerful open-source alternative to closed models like GPT-4o-mini.
According to Mistral AI, this model has recorded an MMLU (Massive Multitask Language Understanding) accuracy of over 81% and demonstrates a speed of 150 tokens per second. Additionally, it maximized processing speed by maintaining fewer layers than existing models. This allows it to achieve a speed more than three times faster than Llama 3.3 70B in the same hardware environment.
Mistral Small 3 can be utilized in various applications, including ▲ conversational AI requiring fast responses, ▲ low-latency function calls, ▲ building domain-specific expert models, and ▲ local deployment environments. It can be particularly applied in various industries such as finance, healthcare, and manufacturing.
According to the company, financial services corporations can use Mistral Small 3 for fraud detection. In the healthcare field, it can be used for triage and consultation services. Additionally, in robotics, automotive, and manufacturing, it can be applied for on-device command and control purposes.
Mistral AI stated that individual and corporate users can also run Mistral Small 3 in a local environment. When the model is quantized, it can be smoothly operated on an NVIDIA RTX 4090 graphics card or a MacBook equipped with 32GB of RAM.
Mistral Small 3 is available in Mistral AI's own platform "la Plateforme" under the name "mistral-small-latest" or "mistral-small-2501."