Kimi K3 on DGX Station GB300: VRAM, speed and fit
Kimi K3 at 1.6-bit runs, spilling into system RAM, on the DGX Station GB300 with 1M of context at about 20 tokens a second.
Text generation · Chat · Ollama / LM Studio
Spills to RAM · on DGX Station GB300 · 1.6-bit (IQ1_S) · Ollama / LM Studio · Chat
323 GB, slower
- GPU memory
- 239 GB
- of 252 GB
- Context
- 1M
- auto
- Speed
- 20 tok/s
- Fast
- First reply
- 5.4 s
- 2K prompt
Where the memory goes
- Model
- 534 GB
- Context
- 27.4 GB
- Overhead
- 1.0 GB
- OS
- 0.0 GB
- Free
- 0.0 GB
- In RAM
- 323 GB
- Memory
- 239 GB of 252 GB · 323 GB in system RAM
- Context
- 1M tokens · auto
- Speed
- 20 tok/s · fast
- First reply
- 5.4 s for a 2K prompt
- Running cost
- $23/mo at home · $1,162/mo on a cloud 4× MI300X
Validation: Validated with conditions
| Check | Target | Result | Status |
|---|---|---|---|
| Fits in GPU memory | 252 GB usable | 562 GB, 323 GB in RAM | Check |
| Headroom | 5% free | GPU full, rest in RAM | Check |
| Context for chat | 16K or more | 1M | Pass |
| Writing speed | 15 tok/s or more | 20 tok/s | Pass |
| First reply | under 3 s, 2K prompt | 5.4 s | Check |
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
- Kimi K3
- Parameters
- 2.8T, 104B active
- Precision
- 1.6-bit (IQ1_S)
- Context
- 1M, auto
- Chats at once
- 1
- Use
- Chat
- Runs in
- Ollama / LM Studio
- Hardware
- DGX Station GB300
- GPU can use
- 252 GB of 252 GB
- Spill space
- 480 GB LPDDR5X on the Grace CPU
Worth trying
- 8× RTX PRO 6000 Blackwell: Runs it with 1M context
- 8× H200 NVL: Runs it with 1M context
Running cost
- At home
- $23/mo
- 1040 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/Kimi-K3-GGUF- vLLM
vllm serve moonshotai/Kimi-K3- Model card
- huggingface.co/moonshotai/Kimi-K3
Kimi K3 on similar machines
- Ryzen AI Max+ 395 128 GBToo big
- RTX SparkToo big
- DGX SparkToo big
- Dual DGX SparkToo big
- 2× RTX PRO 6000 BlackwellToo big
- 4× RTX PRO 6000 BlackwellToo big
- 8× RTX PRO 6000 BlackwellRuns
- A100 80 GB SXMToo big
All machines: Kimi K3 VRAM requirements. Everything that runs here: what can the DGX Station GB300 run?
Other models that run on DGX Station GB300
Questions
Can Kimi K3 run on DGX Station GB300?
Kimi K3 at 1.6-bit runs, spilling into system RAM, on the DGX Station GB300 with 1M of context at about 20 tokens a second.
Report VR-0776865, calculated Nov 22, 2026 from published specs, not benchmarked.