Mistral AI, a French unicorn, launched its Mistral3 series of models on December 2nd, including three small dense models: 3B, 8B, and 14B, as well as the strongest Mistral Large3 to date. The series covers all scenarios from edge devices to enterprise-level inference. The new models continue to use the Apache 2.0 open-source license, with weights already uploaded to Hugging Face and GitHub, allowing commercial free use.
The Mistral3 series maintains the low latency advantage of its predecessors while expanding the context length to 128K. It performs on par or slightly better than the Llama3.1 equivalent model in mainstream benchmarks such as MMLU, HumanEval, and MT-Bench. The company stated that through a hybrid design of "sliding window attention + grouped query attention," the 14B version can complete full context reasoning of 128K on a single A100, with a 42% increase in throughput in batch scenarios, offering higher cost-effectiveness for applications such as academic research, business analysis, and educational content generation.
Image source note: The image is AI-generated, and the image licensing service provider is Midjourney.
