DeepSeek V4 Flash 284B
DeepSeek V4 Flash 284B needs roughly 194.5 GB VRAM at Q4_K_M quantization (637.0 GB at FP16). 7 GPUs we track can run it fully in VRAM at 8k context.
7 GPUs run this natively · 1 with CPU offload
DeepSeek V4 Flash 284B is a Mixture of Experts (MoE) model with 284B total parameters but only 13B active per token developed by DeepSeek. April 2026 lighter sibling of DeepSeek V4 Pro — 284B total parameters with 13B active, same 1M-token context and hybrid CSA/HCA attention design as Pro, but at a fraction of the total weight count. MIT licensed, with text, vision, and video inputs.
To run DeepSeek V4 Flash 284B locally: Q2_K needs roughly 100-105GB — still multi-GPU or a high-memory unified system (128GB+ Mac Studio), but a large step down from Pro's 400-500GB. Not a single consumer-GPU model at any quantization. As a MoE model, inference speed depends on active parameters (13B) rather than total size.
MMLU-Pro 86.3 is within striking distance of the full V4 Pro (87.5) despite roughly a fifth of the total parameters, which is the point of the Flash variant — most of Pro's quality at a size that's actually reachable outside a datacenter.
VRAM at each quantization
DeepSeek V4 Flash 284B natively supports a longer context window, but the table below is capped at 8k for comparability — KV cache grows linearly with context length.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP32 | 1136.0 GB | 0.72 GB | 1273.1 GB |
| BF16 | 568.0 GB | 0.72 GB | 637.0 GB |
| FP16 | 568.0 GB | 0.72 GB | 637.0 GB |
| Q8_0 | 301.9 GB | 0.72 GB | 338.9 GB |
| Q6_K | 233.2 GB | 0.72 GB | 261.9 GB |
| Q5_K_M | 202.2 GB | 0.72 GB | 227.3 GB |
| Q4_K_M | 173.0 GB | 0.72 GB | 194.5 GB |
| Q3_K_M | 136.6 GB | 0.72 GB | 153.8 GB |
| Q2_Krec | 108.2 GB | 0.72 GB | 122.0 GB |
| NVFP4cuda | 142.0 GB | 0.72 GB | 159.8 GB |
KV cache is calculated at 8k context (FP16). Note that NVFP4 only runs on CUDA GPUs. Turn on TurboQuant in the calculator above for lower KV cache estimates.
Benchmarks
GPUs that run DeepSeek V4 Flash 284B natively (7)
- AMD Instinct MI300XQ3_K_M · 159.8 t/s
- Apple M4 Ultra (384GB)Q8_0 · 18.7 t/s
- Apple M4 Ultra (192GB)Q3_K_M · 40.5 t/s
- Apple M3 Ultra (512GB)Q8_0 · 14 t/s
- Apple M3 Ultra (256GB)Q5_K_M · 20.8 t/s
- Apple M2 Ultra (384GB)Q8_0 · 13.7 t/s
- Apple M2 Ultra (192GB)Q3_K_M · 29.7 t/s
Plus 1 GPUs that run it with CPU offload (slower)
- Intel Data Center GPU Max 1550Q2_K · 123.6 t/s
Notes
Lighter V4 sibling — 284B total / 13B active, same 1M context and hybrid CSA/HCA attention as Pro.
Frequently asked questions
- What are the VRAM requirements for DeepSeek V4 Flash 284B?
- DeepSeek V4 Flash 284B requires approximately 194.5 GB of VRAM at Q4_K_M quantization, 338.9 GB at Q8, and 637.0 GB at FP16. These numbers assume 8k context window; VRAM scales linearly with context length due to the KV cache.
- How many parameters does DeepSeek V4 Flash 284B have?
- DeepSeek V4 Flash 284B has 284 billion total parameters, but only 13 billion are active per token thanks to its Mixture of Experts (MoE) architecture. This makes inference significantly faster than the total parameter count suggests.
- How capable is DeepSeek V4 Flash 284B?
- DeepSeek V4 Flash 284B achieves an MMLU-Pro score of 86.3, placing it among the most capable open-weight models available — competitive with frontier systems on general knowledge and reasoning.
- Can DeepSeek V4 Flash 284B run on a 16 GB GPU?
- No. At Q4_K_M, DeepSeek V4 Flash 284B needs 194.5 GB of VRAM — more than 16 GB. You will need a multi-GPU server.
- Can DeepSeek V4 Flash 284B run on a 24 GB GPU?
- No. Even at Q4_K_M, DeepSeek V4 Flash 284B needs 194.5 GB. Consider a multi-GPU server with 80 GB+ total VRAM.
- What is the smallest quantization for DeepSeek V4 Flash 284B that fits in 24 GB of VRAM?
- DeepSeek V4 Flash 284B cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 122.0 GB at Q2_K.
- What GPU do I need to run DeepSeek V4 Flash 284B locally?
- You need a multi-GPU server. At Q4_K_M, DeepSeek V4 Flash 284B needs 194.5 GB VRAM, more than any single consumer GPU. Consider 2–4× H100 or A100 GPUs.