DeepSeek R1 Zero
Zero-eval
by DeepSeek
+
+
+
+
About
DeepSeek-R1-Zero was introduced as an experimental variant trained with minimal human supervision, designed to develop reasoning patterns through self-guided reinforcement learning. Built to explore how models can discover analytical strategies independently, it represents research into autonomous reasoning capability development.
+
+
+
+
Timeline
AnnouncedJan 20, 2025
ReleasedJan 20, 2025
+
+
+
+
Specifications
Training Tokens14.8T
+
+
+
+
License & Family
License
MIT
Base ModelDeepSeek-V3
Performance Overview
Performance metrics and category breakdown
Overall Performance
4 benchmarks
Average Score
76.5%
Best Score
95.9%
High Performers (80%+)
2+
+
+
+
All Benchmark Results for DeepSeek R1 Zero
Complete list of benchmark scores with detailed information
| MATH-500 | text | 0.96 | 95.9% | Self-reported | |
| AIME 2024 | text | 0.87 | 86.7% | Self-reported | |
| GPQA | text | 0.73 | 73.3% | Self-reported | |
| LiveCodeBench | text | 0.50 | 50.0% | Self-reported |
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+