LLMDeepSeek
deepseek-ai/DeepSeek-R1
deepseek-ai/DeepSeek-R1
Reasoning-focused MoE LLM, 671B total with 37B active. State-of-the-art chain-of-thought reasoning.
Parameters
38.5B
Context
160K
Downloads
8M
Likes
13.6K
Architecture
deepseek_v3
61 layers · 128 KV heads · head 56 · hidden 7168
GPU Requirements
VRAM / RAM / disk estimates by quantization and usage scenario
| Quantization | Minimum | Recommended | Production |
|---|---|---|---|
| FP1616bit | 1586.0 GB RAM 2379.0 · Disk 1988.7 | 1673.6 GB RAM 2510.4 · Disk 2014.7 | 1842.4 GB RAM 2763.6 · Disk 2053.7 |
| FP88bit | 795.5 GB RAM 1193.2 · Disk 1013.9 | 847.2 GB RAM 1270.7 · Disk 1039.9 | 980.0 GB RAM 1470.0 · Disk 1078.9 |
| INT88bit | 795.5 GB RAM 1193.2 · Disk 1013.9 | 847.2 GB RAM 1270.7 · Disk 1039.9 | 980.0 GB RAM 1470.0 · Disk 1078.9 |
| AWQ4bit | 415.0 GB RAM 622.6 · Disk 544.7 | 449.4 GB RAM 674.1 · Disk 570.7 | 565.0 GB RAM 847.4 · Disk 609.7 |
| GPTQ4bit | 420.0 GB RAM 630.0 · Disk 550.8 | 454.6 GB RAM 681.9 · Disk 576.8 | 570.4 GB RAM 855.5 · Disk 615.8 |
| GGUF4bit | 429.9 GB RAM 644.8 · Disk 563.0 | 464.9 GB RAM 697.4 · Disk 589.0 | 581.1 GB RAM 871.7 · Disk 628.0 |
GPU Recommendations
Best ValueBEST
8× H200 SXM
1128 GB VRAM
$20,440/mo
≈ 47.4 tok/s
Effective production throughput ≈ 33.2 tok/s
score 0.326
Cheapest
8× A100 80GB
640 GB VRAM
$7,008/mo
≈ 20.1 tok/s
Effective production throughput ≈ 14.1 tok/s
score 0.4479
Performance
4× H200 SXM
564 GB VRAM
$10,220/mo
≈ 35.5 tok/s
Effective production throughput ≈ 24.8 tok/s
score 0.4569
Min Complexity
4× H200 SXM
564 GB VRAM
$10,220/mo
≈ 35.5 tok/s
Effective production throughput ≈ 24.8 tok/s
score 0.4569
💬 Community — real deployment experience
0 Articles · 0 Benchmarks