vanchin/deepseek-r1

DeepSeek-R1 is a 671B-parameter Mixture of Experts (MoE) reasoning model open-sourced by DeepSeek in January 2025, activating only 37B parameters during inference. It is the first 100B+ parameter model trained purely via reinforcement learning without supervised fine-tuning, enabling the natural emergence of Chain-of-Thought (CoT). By incorporating cold-start data prior to RL, the model resolves the repetition and language mixing issues of R1-Zero, achieving performance on par with OpenAI o1 across math, coding, and reasoning tasks.