Skip to content

DeepSeek V4 Pro on 8× H200 NVL: VRAM, speed and fit

DeepSeek V4 Pro at FP4 + FP8 runs on the 8× H200 NVL with 1M of context at about 64 tokens a second.

Text generation · Chat · Ollama / LM Studio

Runs comfortably · on 8× H200 NVL · FP4 + FP8 · Ollama / LM Studio · Chat

GPU memory
820 GB
of 1,128 GB
Context
1M
auto
Speed
64 tok/s
Very fast
First reply
0.4 s
2K prompt

Where the memory goes

Model
805 GB
Context
8.7 GB
Overhead
5.7 GB
OS
0.0 GB
Free
308 GB
Memory
820 GB of 1,128 GB
Context
1M tokens · auto
Speed
64 tok/s · very fast
First reply
0.4 s for a 2K prompt
Running cost
$65/mo at home · $2,325/mo on a cloud 8× MI300X

Validation: Validated

What DeepSeek V4 Pro on 8× H200 NVL was checked against
CheckTargetResultStatus
Fits in GPU memory1,128 GB usable820 GB neededPass
Headroom5% free308 GB free (27%)Pass
Context for chat16K or more1MPass
Writing speed15 tok/s or more64 tok/sPass
First replyunder 3 s, 2K prompt0.4 sPass

Targets: a reply you can read along with, a first reply the use case tolerates, and 5% of memory left free. Calculated from published specs, not benchmarked.

The setup

Model
DeepSeek V4 Pro
Parameters
1.6T, 49B active
Precision
FP4 + FP8
Context
1M, auto
Chats at once
1
Use
Chat
Runs in
Ollama / LM Studio
Hardware
8× H200 NVL
GPU can use
1,128 GB of 1,128 GB

Running cost

At home
$65/mo
2960 W at the wall
Renting instead
$2,325/mo
8× MI300X at $19/h

4 h a day, power at $0.18 per kWh, cloud at RunPod on-demand prices as of Nov 22, 2026.

How to run it

llama.cpp
llama-server -hf unsloth/DeepSeek-V4-Pro-0813-GGUF
vLLM
vllm serve deepseek-ai/DeepSeek-V4-Pro
MLX (Mac)
mlx_lm.chat --model mlx-community/DeepSeek-V4-Pro-4bit

DeepSeek V4 Pro on similar machines

All machines: DeepSeek V4 Pro VRAM requirements. Everything that runs here: what can the 8× H200 NVL run?

Other models that run on 8× H200 NVL

Questions

Can DeepSeek V4 Pro run on 8× H200 NVL?

DeepSeek V4 Pro at FP4 + FP8 runs on the 8× H200 NVL with 1M of context at about 64 tokens a second.

Report VR-1QNFWDX, calculated Nov 22, 2026 from published specs, not benchmarked.