DeepSeek-r1:7b

DeepSeek-r1:7b is the 7B parameter distilled reasoning model (Qwen-based) that leverages 800K samples from the 671B DeepSeek-R1 parent, delivering top-tier performance on complex benchmarks like MATH-500 (92.8 score).

This is the DeepSeek-R1-Distill-Qwen-7B checkpoint: a highly capable, 7-billion-parameter model optimized for reasoning tasks. It is a powerful distillation, fine-tuned with over 800,000 samples generated by the flagship 671B DeepSeek-R1 model (Qwen-based). This process transfers advanced reasoning capabilities, resulting in state-of-the-art performance for its size class, notably achieving a 92.8 on the MATH-500 benchmark and a 55.5 on AIME 2024 (pass@1). Deployed easily via tools like Ollama, this MIT-licensed model provides accessible, commercial-use reasoning power for mathematics, coding, and complex problem-solving on consumer-grade hardware.

https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B