Inference model DeepSeek-R1-Lite preview goes live, demystifying the o1 inference process

DeepSeek-R1-Lite is trained using reinforcement learning, and the reasoning process includes a lot of reflection and verification, supporting thought chains up to tens of thousands of words long; in complex tasks such as math and programming, DeepSeek-R1-Lite surpasses GPT-4o in AMC, Codeforces, and other reviews, demonstrating excellent results; reasoning efficiency is positively correlated with thought chain length , compared with traditional voting methods, long chain reasoning improves accuracy and efficiency.