DeepSeek V4 Flash
deepseek-ai/DeepSeek-V4-Flash
Parameters
284B
Context
1024K
Downloads
2.3M
Likes
2.1K
Architecture
moe_transformer
GPU Requirements
VRAM / RAM / disk estimates by quantization and usage scenario
| Quantization | Minimum | Recommended | Production |
|---|---|---|---|
| FP1616bit | 674.1 GB RAM 1011.2 · Disk 864.2 | 720.3 GB RAM 1080.4 · Disk 890.2 | 847.6 GB RAM 1271.4 · Disk 929.2 |
| FP88bit | 339.5 GB RAM 509.3 · Disk 451.6 | 370.5 GB RAM 555.7 · Disk 477.6 | 482.6 GB RAM 723.9 · Disk 516.6 |
| INT88bit | 339.5 GB RAM 509.3 · Disk 451.6 | 370.5 GB RAM 555.7 · Disk 477.6 | 482.6 GB RAM 723.9 · Disk 516.6 |
| AWQ4bit | 178.5 GB RAM 267.8 · Disk 253.0 | 202.2 GB RAM 303.2 · Disk 279.0 | 307.0 GB RAM 460.4 · Disk 318.0 |
| GPTQ4bit | 180.6 GB RAM 270.9 · Disk 255.6 | 204.3 GB RAM 306.5 · Disk 281.6 | 309.2 GB RAM 463.8 · Disk 320.6 |
| GGUF4bit | 184.8 GB RAM 277.2 · Disk 260.8 | 208.7 GB RAM 313.1 · Disk 286.8 | 313.8 GB RAM 470.7 · Disk 325.8 |
GPU Recommendations
Best ValueBEST
8× H200 SXM
1128 GB VRAM
$20,440/mo
≈ 111.9 tok/s
Effective production throughput ≈ 78.3 tok/s
score 0.2675
Cheapest
8× RTX A6000
384 GB VRAM
$1,678/mo
≈ 17.9 tok/s
Effective production throughput ≈ 12.5 tok/s
score 0.483
Performance
4× H200 SXM
564 GB VRAM
$10,220/mo
≈ 84 tok/s
Effective production throughput ≈ 58.8 tok/s
score 0.365
Min Complexity
2× H200 SXM
282 GB VRAM
$5,110/mo
≈ 56 tok/s
Effective production throughput ≈ 39.2 tok/s
score 0.5871
💬 Community — real deployment experience
0 Articles · 0 Benchmarks