DeepSeek-R1-Distill-Qwen-1.5B

DeepSeek-R1-Distill-Qwen-1.5B is the smallest member of the DeepSeek-R1 distilled model family, released in January 2025 alongside the R1 reasoning model. It is fine-tuned from Qwen2.5-1.5B on 800,000 reasoning samples curated with DeepSeek-R1, giving it explicit chain-of-thought behavior: it opens a <think> block, reasons step by step, then answers.

The model is notable for bringing reasoning capability to phones and edge hardware. Community conversions have shrunk it to around 1GB (int4), and it runs in on-device runtimes at over 100 tokens per second on modern phone chips. The weights carry the MIT License, with the underlying Qwen2.5 base under Apache 2.0. DeepSeek recommends a temperature of 0.5-0.7, no system prompt, and an explicit step-by-step directive for math tasks.