Skip to content

DeepSeek V3.2 on 4× H200 NVL: VRAM, speed and fit

DeepSeek V3.2 at 6-bit runs on the 4× H200 NVL with 160K of context at about 79 tokens a second.

Text generation · Chat · Ollama / LM Studio

Runs comfortably · on 4× H200 NVL · 6-bit (Q6_K) · Ollama / LM Studio · Chat

GPU memory
527 GB
of 564 GB
Context
160K
auto
Speed
79 tok/s
61 when full
First reply
0.6 s
2K prompt

Where the memory goes

Model
512 GB
Context
11.9 GB
Overhead
3.0 GB
OS
0.0 GB
Free
36.7 GB
Memory
527 GB of 564 GB
Context
160K tokens · auto
Speed
79 tok/s · 61 when full
First reply
0.6 s for a 2K prompt
Running cost
$33/mo at home · $1,162/mo on a cloud 4× MI300X

Validation: Validated

What DeepSeek V3.2 on 4× H200 NVL was checked against
CheckTargetResultStatus
Fits in GPU memory564 GB usable527 GB neededPass
Headroom5% free36.7 GB free (7%)Pass
Context for chat16K or more160KPass
Writing speed15 tok/s or more79 tok/sPass
First replyunder 3 s, 2K prompt0.6 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 V3.2
Parameters
671B, 37B active
Precision
6-bit (Q6_K)
Context
160K, auto
Chats at once
1
Use
Chat
Runs in
Ollama / LM Studio
Hardware
4× H200 NVL
GPU can use
564 GB of 564 GB

Running cost

At home
$33/mo
1520 W at the wall
Renting instead
$1,162/mo
4× MI300X at $9.56/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-V3.2-GGUF
vLLM
vllm serve deepseek-ai/DeepSeek-V3.2
MLX (Mac)
mlx_lm.chat --model mlx-community/DeepSeek-V3.2-4bit

DeepSeek V3.2 on similar machines

All machines: DeepSeek V3.2 VRAM requirements. Everything that runs here: what can the 4× H200 NVL run?

Other models that run on 4× H200 NVL

Questions

Can DeepSeek V3.2 run on 4× H200 NVL?

DeepSeek V3.2 at 6-bit runs on the 4× H200 NVL with 160K of context at about 79 tokens a second.

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