Mistral, an artificial intelligence (AI) company based in Paris, unveiled its latest model, the Mistral Small 3, on Thursday. The firm, recognized for its open-source large language models (LLMs), has made this new AI model accessible on platforms including Hugging Face. The company asserts that the model has been optimized for processing speed, efficiency, and overall performance, claiming it outperforms larger models. Internal evaluations indicated that Mistral Small 3 surpasses the performance of OpenAI’s GPT-4o mini.
Launch of Mistral Small 3 AI Model
In a recent announcement, Mistral provided insights into the new AI model. The Mistral Small 3 is a latency-optimized model featuring 24 billion parameters. It is offered in both a pre-trained version and an instruction-tuned checkpoint, accommodating various use cases. The model is licensed under the Apache 2.0 license, allowing for both academic and commercial applications. Mistral has transitioned from its previous Mistral Research License (MRL), which permitted only academic and research use.
The company clarified that the model has not undergone reinforcement learning (RL) training nor has it utilized synthetic data generated from other AI models or digital sources.
Based on its internal assessments, Mistral concluded that the Small 3 model exceeds the latency performance of GPT-4o mini. Additionally, it performed favorably against OpenAI’s model in key benchmarks such as the Massive Multitask Language Understanding (MMLU) Pro and the Graduate-Level Google-Proof Q&A (GPQA). The developers noted that despite being three times smaller, this model remains competitive with Llama 3.3 70B.
The company identified several scenarios where the model’s efficiency and speed could be essential for developers. Potential use cases include scenarios requiring rapid-response conversational support, low-latency function calls, or the development of a chatbot with expertise in specific topics by fine-tuning the LLM.
Moreover, the AI model is suitable for organizations that prioritize local inference to protect sensitive or proprietary information. Notably, Mistral Small 3 can operate privately on a single Nvidia RTX 4090 GPU. Developers can access the model via its listing on Hugging Face here.
