On Thursday, Mistral, the Paris-based artificial intelligence (AI) company, unveiled its latest offering, the Mistral Small 3 AI model. Renowned for developing open-source large language models (LLMs), the firm has also made this new model accessible on Hugging Face and various other platforms. Mistral asserts that the Mistral Small 3 was engineered with a focus on processing speed, efficiency, and high performance, claiming it can surpass models that are twice its size. According to the company’s internal evaluations, the model demonstrates superior performance compared to OpenAI’s GPT-4o mini.
Introduction of the Mistral Small 3 AI Model
In an official announcement, the French AI company provided insights into the specifications of the new model. The Mistral Small 3 features 24 billion parameters and is designed to optimize latency. The LLM is being launched with both pre-trained and instruction-tuned checkpoints, allowing for a diverse range of applications. It is released under the Apache 2.0 license, permitting both academic and commercial use, marking a shift away from the previous Mistral Research Licence (MRL), which restricted usage to research-related activities.
The company clarified that the new AI model does not incorporate reinforcement learning (RL) techniques, nor does it utilize synthetic data in its training process.
Based on the company’s internal tests, Mistral Small 3 reportedly surpasses GPT-4o mini concerning latency and achieves higher scores than OpenAI’s model on key benchmarks such as the Massive Multitask Language Understanding (MMLU) Pro and Graduate-Level Google-Proof Q&A (GPQA). Developers noted that the model is competitive with the larger Llama 3.3 70B model, despite being three times smaller in size.
Mistral indicated that this model caters to use cases where speed and efficiency are crucial for developers. Suggested applications include contexts requiring immediate response for conversational assistance, low-latency function calling, and scenarios where developers aim to create specialized chatbots by fine-tuning the LLM.
The Mistral Small 3 model is also suited for organizations prioritizing local inference to protect sensitive or proprietary information. Importantly, it can be operated privately on a single Nvidia RTX 4090 GPU. Developers can access the model via its Hugging Face listing.
