Model
Model explorer

DeepSeek-R1-Zero

OPEN
DeepSeek · DeepSeek-R1 family · released Jan 20, 2025

Trained by pure large-scale RL on V3-Base with no SFT cold start; first open proof reasoning can emerge from RL alone, but suffers repetition/readability/language-mixing issues.

ReasoningCodingVisionFunction callingTool useAgentic
1435.8
Elo · rank #133
Parameters
671B
Active params
37B (MoE)
Context
128K tokens
Architecture
DeepSeekMoE: 256 routed + 1 shared expert, top-8 routing, 61 layers (3 dense + 58 MoE), Multi-head Latent Attention (MLA)
License
MIT
Languages
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
AIMEMath71.0%#101
best: GPT-5.2 · 100.0%
ARC-AGI-1Reasoning14.0%#17
best: GPT-5.6 · 96.5%
CodeforcesCoding1444#36
best: DeepSeek-V4-Pro (Think Max) · 3206
GPQA DiamondReasoning73.3%#108
best: GPT-5.6 · 94.6%
LiveCodeBenchCoding50.0%#118
best: DeepSeek-V4-Pro (Think Max) · 93.5%
MATH-500Math95.9%#29
best: GPT-5 · 99.4%
Run it locally
VRAM @ Q4
404 GB
VRAM @ FP16
1342 GB
Fits on (Q4)
Multi-node cluster required
Same 671B MoE footprint as DeepSeek-R1; does not fit a single 24GB 4090 even at Q4, and community quantization interest has been lower than for the instruction-tuned R1.
Quantizations
GGUF Q4_K_M · GGUF Q8_0
API price weights · each benchmark row carries its own source badge (see methodology)