vLLM on the NVIDIA DGX Spark
On the NVIDIA DGX Spark, vLLM runs DeepSeek V4 Flash 162B · 13B active at 4-bit at a median 13.9 tokens per second for generation, from 1 plain run (200K context).
Speed by model on the NVIDIA DGX Spark
Plain runs only: one device, one request, no speculative decoding, the whole model in memory. Each row is one model size at one bit class, with the other engines' plain runs of the same model on this card beside it. Context length and build still differ between runs, and each side shows its own.
| Model | vLLM | Other engines |
|---|---|---|
| DeepSeek V4 Flash 162B · 13B active 4-bit | 13.9 t/sone report, 200K context | no other engine reported |
| Gemma 4 26B · 4B active 4-bit | 48.2 t/sone report, context not stated | llama.cpp 69.9 t/sone report, 256K context |
| Qwen3.6 35B · 3B active 4-bit | 78.3 t/sone report, 128K context | no other engine reported |
36 vLLM reports on the NVIDIA DGX Spark
Newest first. A run on several cards, with speculative decoding, with part of the model in system RAM or with batched requests says so, since each describes a different setup.
- Qwen3.8 125B · 6B active · NVFP486.0 t/s(2 cards, speculative decoding)
- Qwen3.8 · NVFP4 / MXFP8 mixed118.6 t/s(2 cards, speculative decoding)
- Qwen3.8 · NVFP4 / MXFP8 mixed78.6 t/s(speculative decoding)
- Gemma 4 26B · 4B active · NVFP448.2 t/s
- Qwen3.8 125B · 6B active · FP839.8 t/s(2 cards, speculative decoding)
- DeepSeek V4.1 Flash 552B · 16B active · MXFP438.0 t/s(4 cards, speculative decoding)
- DeepSeek V4 Flash 162B · 13B active · NVFP413.9 t/s
- DeepSeek V4 Flash 284B · 13B active · 2-bit expert planes9.8 t/s(speculative decoding)
- DeepSeek V4 Flash 285B · 13B active · NVFP480.1 t/s(2 cards, speculative decoding)
- DeepSeek V4 Flash 284B · 13B active · FP841.0 t/s(2 cards, part in system RAM)
- Qwen3.8 125B · 6B active · NVFP430.0 t/s(speculative decoding)
- Ornith1.5 35B · 3B active · NVFP479.0 t/s(speculative decoding)
- GLM-5.3 320B · 18B active · NVFP440.0 t/s(2 cards, speculative decoding)
- Kimi K3 2800B · 104B active20.0 t/s(16 cards)
- Qwen3.6 35B · 3B active · NVFP478.3 t/s
- Mimo 2.6 309B · 15B active · MXFP453.3 t/s(2 cards, speculative decoding)
- GLM-5.3 320B · 18B active · NVFP469.3 t/s(4 cards, speculative decoding)
- GLM-5.3 320B · 18B active · NVFP430.1 t/s(2 cards, speculative decoding)
- Qwen3.8 125B · 6B active · NVFP443.9 t/s(speculative decoding, part in system RAM)
- Qwen3.8 125B · 6B active · NVFP427.8 t/s(speculative decoding)
- DeepSeek V4 Flash 284B · 13B active72.4 t/s(2 cards, speculative decoding)
- Qwen3.8 125B · 6B active · NVFP434.0 t/s(speculative decoding, part in system RAM)
- Mimo 2.6 1020B · 42B active68.3 t/s(8 cards, speculative decoding)
- Kimi K3 2800B · 104B active30.0 t/s(16 cards, speculative decoding)
- Qwen3.8 125B · 6B active · NVFP443.9 t/s(speculative decoding, part in system RAM)
- Ling-3.0 · INT440.9 t/s(speculative decoding)
- DeepSeek V4 Flash 284B · 13B active · MXFP8 x MXFP449.4 t/s(2 cards, speculative decoding)
- GLM-5.2 · NVFP4PP 512.0 t/s(speculative decoding)
- DeepSeek V4 Flash 284B · 13B active · NVFP420.6 t/s(part in system RAM)
- DeepSeek V4 Flash 284B · 13B active · 2-bit19.1 t/s
Frequently asked
How fast is vLLM on the NVIDIA DGX Spark?
On the NVIDIA DGX Spark, vLLM runs DeepSeek V4 Flash 162B · 13B active at 4-bit at a median 13.9 tokens per second for generation, from 1 plain run (200K context).
Is vLLM faster than other engines on the NVIDIA DGX Spark?
Same card, same model size and bit class, plain runs only: Gemma 4 26B · 4B active at 4-bit gives vLLM 48.2 t/s (1 run) and llama.cpp 69.9 t/s (1). Context length and engine build differ between the runs, so treat it as a lead, not a benchmark.