Model
Model explorer
DeepSeek-R1-0528
OPENDeepSeek · DeepSeek-R1 family · released May 28, 2025
Major R1 refresh with deeper chain-of-thought (AIME24 79.8 to 91.4), fewer hallucinations, and added JSON/function calling; approached o3/Gemini 2.5 Pro.
ReasoningCodingVisionFunction callingTool useAgentic
1757.4
Elo · rank #83
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)
$0.55 / $2.19
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
best: Claude Opus 4.5 · 89.4%
best: GPT-5.2 · 100.0%
best: GPT-5.6 · 96.5%
best: GPT-5.6 · 92.5%
best: Hunyuan-A13B · 78.3%
best: DeepSeek-V4-Pro (Think Max) · 3206
best: GPT-5.6 · 94.6%
best: Claude Sonnet 5 · 57.4%
best: DeepSeek-V4-Pro (Think Max) · 93.5%
best: Claude Fable 5 · 91.5%
best: Qwen3.7-Max · 95.0%
best: GPT-4.5 · 62.5%
best: Claude Fable 5 · 95.0%
best: Claude Opus 4.1 · 82.4%
Run it locally
VRAM @ Q4
404 GB
VRAM @ FP16
1342 GB
Fits on (Q4)
Multi-node cluster required
671B MoE far exceeds a single 24GB 4090; Unsloth's dynamic quants shrink it to ~185GB for CPU-offload rigs, still requiring far more than one consumer GPU.
Quantizations
GGUF Q4_K_M · GGUF IQ1/IQ2 (Unsloth Dynamic) · AWQ · MLX
DeepSeek-R1 family
Elo progression across releases
API price $0.55/$2.19 · each benchmark row carries its own source badge (see methodology)