DeepSeek V4 Flash 284B
DeepSeek V4 Flash 284B needs roughly 159.8GB VRAM at Q4_K_M quantization (637.0GB at FP16). 10 GPUs we track can run it fully in VRAM at 8k context.
DeepSeek284B params13B active (MoE)1024k contextMITCommercial use ok
VRAM at each quantization
Assumes 8k context. KV cache grows linearly with context length.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP16 | 568.0 GB | 0.72 GB | 637.0 GB |
| Q8 | 284.0 GB | 0.72 GB | 318.9 GB |
| Q6_K | 213.0 GB | 0.72 GB | 239.4 GB |
| Q5_K_M | 177.5 GB | 0.72 GB | 199.6 GB |
| Q4_K_M | 142.0 GB | 0.72 GB | 159.8 GB |
| Q3_K_M | 113.6 GB | 0.72 GB | 128.0 GB |
| Q2_K | 85.2 GB | 0.72 GB | 96.2 GB |
Benchmarks
Benchmarks for this model are not yet available on the Open LLM Leaderboard v2. This is common for recently released models. Check back soon.
GPUs that run DeepSeek V4 Flash 284B natively (10)
- NVIDIA DGX Spark (128GB)Q2_K · 77 t/s
- AMD Instinct MI300XQ4_K_M · 896.9 t/s
- AMD Strix Halo (128GB)Q2_K · 72.2 t/s
- Apple M4 Ultra (384GB)Q8 · 92.4 t/s
- Apple M4 Ultra (192GB)Q4_K_M · 184.8 t/s
- Apple M4 Max (128GB)Q2_K · 154 t/s
- Apple M3 Max (128GB)Q2_K · 112.8 t/s
- Apple M2 Ultra (384GB)Q8 · 67.7 t/s
- Apple M2 Ultra (192GB)Q4_K_M · 135.4 t/s
- Apple M1 Ultra (128GB)Q2_K · 225.6 t/s
Plus 2 GPUs that run it with CPU offload (slower)
- NVIDIA H100 80GBQ2_K · 214.7 t/s
- NVIDIA A100 80GBQ2_K · 130.7 t/s
Notes
Lighter V4 sibling — 284B total / 13B active, same 1M context and hybrid CSA/HCA attention as Pro.