DeepSeek V4 Flash VRAM Requirements
DeepSeek V4 Flash needs about 144 GB of VRAM at FP4 + FP8 (official) with an 8K context: 143 GB for weights, 57 MB for the KV cache and 0.8 GB of runtime overhead. At Q4_K_M it needs 161 GB. At Q8_0 it needs 282 GB and at full FP16 precision 530 GB. The smallest GPU that runs it fully is the Apple M2 Ultra 192GB.
144 GB VRAM
DeepSeek V4 Flash at FP4 + FP8 (official) with 8K context and the KV cache in FP16.
- Weights 143 GB
- KV cache 57 MB
- Runtime overhead 0.8 GB
Mixture of experts: all 284B parameters need memory, but only 13B are read per token, so it runs at the speed of a much smaller model.
Which GPUs can run it
- RTX 3060 12GBToo small
- RTX 4060 Ti 16GBToo small
- RTX 3090 24GBToo small
- Radeon RX 7900 XTX 24GBToo small
- RTX 4090 24GBToo small
- RTX 5090 32GBToo small
- RTX 6000 Ada 48GB31% on GPU~8.1 tok/s
- A100 80GB52% on GPU~11 tok/s
- H100 80GB SXM52% on GPU~12 tok/s
- Ryzen AI Max+ 395 128GBToo small
- Apple M4 Max 128GBToo small
- Apple M5 Max 128GBToo small
- RTX PRO 6000 Blackwell 96GB63% on GPU~14 tok/s
- NVIDIA DGX Spark 128GBToo small
- Apple M3 Ultra 512GBFits~57 tok/s
Multi-GPU: about 7× 24 GB cards (RTX 3090 / 4090) with the model split across them.
DeepSeek V4 Flash VRAM by quantization
Total memory with an FP16 KV cache. "Smallest tier" is the smallest standard VRAM size with a 6% safety margin at 8K.
| Quantization | Bits/weight | Weights | Total at 8K | Total at 32K | Smallest tier |
|---|---|---|---|---|---|
| FP4 + FP8 (official) | 4.33 | 143 GB | 144 GB | 144 GB | Multi-GPU / server |
| Q4_K_M | 4.85 | 160 GB | 161 GB | 161 GB | Multi-GPU / server |
| Q5_K_M | 5.69 | 188 GB | 189 GB | 189 GB | Multi-GPU / server |
| Q6_K | 6.59 | 218 GB | 219 GB | 219 GB | Multi-GPU / server |
| Q8_0 | 8.50 | 281 GB | 282 GB | 282 GB | Multi-GPU / server |
| FP16 / BF16 | 16.00 | 529 GB | 530 GB | 530 GB | Multi-GPU / server |
| Q3_K_M | 3.91 | 129 GB | 130 GB | 130 GB | Multi-GPU / server |
| Q2_K | 3.00 | 99.2 GB | 100 GB | 100 GB | 128 GB |
VRAM by context length
At FP4 + FP8 (official). Compressed attention: every layer keeps a 128-token local window, 20 layers also keep one compressed entry per 4 tokens (plus indexer keys) and 20 keep one per 128 tokens. That makes a long context very cheap. These cache figures are our estimate from config.json; real use depends on runtime support.
| Context | KV cache (FP16) | KV cache (Q8_0) | Total, FP16 cache | Total, Q8_0 cache |
|---|---|---|---|---|
| 4K tokens | 31 MB | 16 MB | 144 GB | 144 GB |
| 8K tokens | 57 MB | 30 MB | 144 GB | 144 GB |
| 32K tokens | 0.2 GB | 0.1 GB | 144 GB | 144 GB |
| 128K tokens | 0.8 GB | 0.4 GB | 145 GB | 144 GB |
| 256K tokens | 1.6 GB | 0.9 GB | 145 GB | 145 GB |
GPUs that can run DeepSeek V4 Flash
At FP4 + FP8 (official) with 8K context. "Best quant" is the highest-quality format that fits fully; "max context" is at FP4 + FP8 (official).
| GPU | Usable memory | Status | Best quant | Max context | Speed |
|---|---|---|---|---|---|
| Apple M2 Ultra 192GB | 144 GB | Tight fit | Q3_K_M | – | ~56 tok/s |
| AMD Instinct MI300X 192GB | 192 GB | Fits | FP4 + FP8 (official) | 256K | ~149 tok/s |
| Apple M3 Ultra 512GB | 384 GB | Fits | FP4 + FP8 (official) | 256K | ~57 tok/s |
| NVIDIA H200 141GB | 141 GB | 92% on GPU | Q3_K_M | – | ~51 tok/s |
| RTX PRO 6000 Blackwell 96GB | 96.0 GB | 63% on GPU | – | – | ~14 tok/s |
| A100 80GB | 80.0 GB | 52% on GPU | – | – | ~11 tok/s |
| H100 80GB SXM | 80.0 GB | 52% on GPU | – | – | ~12 tok/s |
| RTX 6000 Ada 48GB | 48.0 GB | 31% on GPU | – | – | ~8.1 tok/s |
| RTX A6000 48GB | 48.0 GB | 31% on GPU | – | – | ~8.1 tok/s |
Multi-GPU: about 7× 24 GB cards (RTX 3090 or 4090) run it at FP4 + FP8 (official), using tensor or layer splitting in llama.cpp, vLLM or ExLlamaV2.
About DeepSeek V4 Flash
The smaller DeepSeek V4: 284B total parameters with about 13B active and a 1M-token context. Experts ship in FP4 and most other weights in FP8, so the official checkpoint is about 143 GB by our estimate. Its compressed attention keeps the cache a small fraction of earlier DeepSeek models. We cap context at 256K here.
| Spec | Value |
|---|---|
| Parameters | 284B total, 13B active (MoE) |
| Layers | 43 |
| Hidden size | 4096 |
| Attention heads / KV heads | 64 / 1 (varies by layer) |
| Head dimension | 512 |
| Vocabulary | 129,280 |
| Max context (used here) | 256K tokens |
| Layers with a growing cache | 40 of 43 |
| KV cache per token (FP16) | 6 KB |
| Released | 2026-04 |
| License | MIT |
DeepSeek V4 Flash VRAM: common questions
Can DeepSeek V4 Flash run on a 192GB or 256GB Mac?
Yes. The native FP4 + FP8 weights are about 143 GB by our estimate, and the compressed cache is small, so a 192 GB machine with a raised GPU memory limit or a 256 GB Mac can hold it.
How much VRAM does DeepSeek V4 Flash need?
About 144 GB at FP4 + FP8 (official) with an 8K context, 282 GB at Q8_0 and 530 GB at FP16. Longer contexts add 0.01 GB per 1,000 tokens with an FP16 cache.
What is the cheapest GPU that can run DeepSeek V4 Flash?
The smallest listed device that fits it fully at FP4 + FP8 (official) with 8K context is the Apple M2 Ultra 192GB (144 GB usable). With 24 GB cards you need about 7 of them.
Does DeepSeek V4 Flash need VRAM for all 284B parameters?
Yes. All experts must be loaded because any token can use any expert. Only 13B parameters are read per token, so it generates text about as fast as a 13B dense model.