Skip to content

Inkling on 4× H200 NVL: VRAM, speed and fit

Inkling at 4-bit runs on the 4× H200 NVL with 4K of context at about 88 tokens a second.

Text generation · Chat · Ollama / LM Studio

Tight fit · on 4× H200 NVL · 4-bit (Q4_K_M) · Ollama / LM Studio · Chat

96% used

GPU memory
542 GB
of 564 GB
Context
4K
auto
Speed
88 tok/s
Very fast
First reply
0.7 s
2K prompt

Where the memory goes

Model
538 GB
Context
1.0 GB
Overhead
3.2 GB
OS
0.0 GB
Free
22.0 GB
Memory
542 GB of 564 GB
Context
4K tokens · auto
Speed
88 tok/s · very fast
First reply
0.7 s for a 2K prompt
Running cost
$33/mo at home · $1,162/mo on a cloud 4× MI300X

Validation: Validated with conditions

What Inkling on 4× H200 NVL was checked against
CheckTargetResultStatus
Fits in GPU memory564 GB usable542 GB neededPass
Headroom5% free22.0 GB free (4%)Check
Context for chat4K or more4KPass
Writing speed15 tok/s or more88 tok/sPass
First replyunder 3 s, 2K prompt0.7 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
Inkling
Parameters
952B, 40.9B active
Precision
4-bit (Q4_K_M)
Context
4K, auto
Chats at once
1
Use
Chat
Runs in
Ollama / LM Studio
Hardware
4× H200 NVL
GPU can use
564 GB of 564 GB

Worth trying

  • 8× RTX PRO 6000 Blackwell: Runs it with 4K context
  • 8× H200 NVL: Runs it with 4K context

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 21, 2026.

How to run it

llama.cpp
llama-server -hf unsloth/inkling-GGUF
vLLM
vllm serve thinkingmachines/Inkling

Inkling on similar machines

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

Other models that run on 4× H200 NVL

Questions

Can Inkling run on 4× H200 NVL?

Inkling at 4-bit runs on the 4× H200 NVL with 4K of context at about 88 tokens a second.

Report VR-15VU1CR, calculated Nov 22, 2026 from published specs, not benchmarked.