DeepSeek-V4-Flash needs about 156 GB of GPU memory for FP8 weights. The smallest fitting setup at 32k context is 1× NVIDIA B200. It supports 3 concurrent conversations with SGLang defaults.
GPU requirements by context length
8k context
GPU
GPUs
Concurrent users
NVIDIA B200
1
11
NVIDIA H200 SXM
2
64
NVIDIA H100 NVL
2
16
NVIDIA RTX PRO 6000 Blackwell Server Edition
2
22
NVIDIA H20
2
22
NVIDIA H100 SXM
4
64
NVIDIA A100 80GB SXM
4
64
NVIDIA L40S
8
62
32k context
GPU
GPUs
Concurrent users
NVIDIA B200
1
3
NVIDIA H200 SXM
2
34
NVIDIA H100 NVL
2
4
NVIDIA RTX PRO 6000 Blackwell Server Edition
2
5
NVIDIA H20
2
5
NVIDIA H100 SXM
4
26
NVIDIA A100 80GB SXM
4
22
NVIDIA L40S
8
16
128k context
GPU
GPUs
Concurrent users
NVIDIA B200
2
17
NVIDIA H200 SXM
2
8
NVIDIA H100 NVL
2
1
NVIDIA RTX PRO 6000 Blackwell Server Edition
2
1
NVIDIA H20
2
1
NVIDIA H100 SXM
4
6
NVIDIA A100 80GB SXM
4
5
NVIDIA L40S
8
4
FP8 weights, SGLang defaults, estimates.
Weights by precision
Weight memory
Precision
Weights
Smallest fitting setup at 32k
FP8
about 156 GB
1x NVIDIA B200
INT4 (AWQ)
about 86 GB
1x NVIDIA B200
Model notes
Attention
Compressed recurrent state
Architecture
Mixture-of-Experts
Layers
43
Hidden size
4,096
Vocabulary
129,280
Native precision
bfloat16
This model keeps a fixed-size recurrent state per conversation instead of a growing token cache, so very long chats cost little extra memory.
Frequently asked questions
Will deepseek-ai/DeepSeek-V4-Flash run on a single H100?
No, a single H100 SXM cannot hold deepseek-ai/DeepSeek-V4-Flash at 32k context with FP8 weights. The model requires multiple GPUs.
What is the cheapest GPU setup for deepseek-ai/DeepSeek-V4-Flash?
The cheapest fitting setup at 32k context is 1x NVIDIA B200 with FP8 weights and SGLang defaults.
How much VRAM does deepseek-ai/DeepSeek-V4-Flash need at 128k context?
At 128k context on the cheapest fitting setup, deepseek-ai/DeepSeek-V4-Flash uses about 156 GB of VRAM per GPU for FP8 weights and about 4.6 GB per conversation for the KV cache. The total VRAM needed depends on the GPU count and parallelism configuration.
Can I run deepseek-ai/DeepSeek-V4-Flash with INT4 quantization?
Yes, deepseek-ai/DeepSeek-V4-Flash has INT4 weight figures. Its weights take about 86 GB in INT4 versus about 156 GB in FP8. INT4 can make the model fit on fewer GPUs, but check inference quality for your workload.
Want a different context length, GPU, or quantization? The calculator runs the same numbers live, preloaded with this model.