DeepSeek-R1-Zero

DeepSeek-R1-Zero, released January 20, 2025, was the precursor to DeepSeek-R1 and the first open-source demonstration that sophisticated reasoning behaviors could emerge from reinforcement learning alone. Trained with Group Relative Policy Optimization (GRPO) directly on the DeepSeek-V3 base model, it skipped supervised fine-tuning entirely and spontaneously developed self-verification, reflection, and long chain-of-thought reasoning.

The trade-off was polish: R1-Zero was prone to repetition, language mixing, and unwieldy formatting, which is why DeepSeek added a cold-start SFT phase to produce the final R1. The experiment was released under the MIT license with full weights and a detailed technical report, and its scientific clarity helped trigger the industry-wide shift toward RL-first reasoning training.