

DeepSeek-R1 Zero
#17 in Reasoning-ModelleDeepSeek · siet 20. Januar 2025 · 2× · tolest 18. Nov. 2026
23
Momentum
DeepSeek-R1-Zero is a reasoning model trained via large-scale reinforcement learning on the DeepSeek-V3-Base model without supervised fine-tuning (SFT) as a preliminary step. It is a Mixture-of-Experts model with 671B total and 37B activated parameters, released as a research model; it shows issues such as endless repetition, poor readability and language mixing.
Momentum-Verloop
24.08.22.11.
Features
| Key Benchmark (%) | AIME 2024 pass@1: 71.0% (86.7% with majority voting) |
| Context Window (Tokens) | 128K tokens; max output 32,768 tokens |
| License | MIT |
| Platform | Open weights on Hugging Face (self-hostable); no official hosted API for R1-Zero |
| Release Date | January 20, 2025 |