Mistral, the French artificial intelligence company, has unveiled their LLM model called Mistral Small 3. This open-source model, with a size of 24B, boasts capabilities close to the GPT-4o-mini, a closed model.
The design approach for Mistral Small 3 focuses on reducing the model’s layers, resulting in low latency and responsiveness of up to 150 tokens per second. Testing has shown that Mistral Small 3 outperforms larger models like Gemma2-27B and Qwen-2.5-32B, approaching the performance level of GPT-4o-mini.
Mistral recommends using this model for tasks requiring speed, such as instant response assistants or as a base model for fine-tuning specialized knowledge. It is also suitable for running models in an organization’s data center. Additionally, the model is licensed under Apache 2.0, allowing for pre-training with DeepSeek-R1 capabilities.
The model can be used on Mistral’s Le Platforme or other platforms like Together.AI, IBM watsonx, or Ollama.
TLDR: Mistral introduces the Mistral Small 3 LLM model with impressive speed and performance capabilities, recommended for tasks requiring quick responses and fine-tuning of specialized knowledge.
Leave a Comment