← All models

DeepSeek-R1-Zero

DeepSeek · 2025-01-20 · 684.5B parameters

DeepSeek-R1-Zero is the reinforcement-learning-only reasoning checkpoint DeepSeek released alongside DeepSeek-R1 on January 20, 2025: GRPO applied directly to DeepSeek-V3-Base with rule-based accuracy and format rewards and no supervised fine-tuning stage. It shares R1's architecture, and its downloadable checkpoint contains exactly 684,489,845,504 parameters including one multi-token-prediction module, which DeepSeek rounds to 671B total and 37B activated. DeepSeek reports poor readability and language mixing in its outputs, which R1's cold-start and later stages were designed to fix.

Benchmark scores