vLLM on the NVIDIA RTX Pro 6000 Blackwell
On the NVIDIA RTX Pro 6000 Blackwell, vLLM runs Gemma 4 31B at 4-bit at a median 51.0 tokens per second for generation, from 1 plain run (16K context).
Speed by model on the NVIDIA RTX Pro 6000 Blackwell
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 |
|---|---|---|
| Gemma 4 31B 4-bit | 51.0 t/sone report, 16K context | no other engine reported |
13 vLLM reports on the NVIDIA RTX Pro 6000 Blackwell
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 8B86.0 t/s(32 requests at once)
- Minimax M3 428B · 23B active · MXFP4113.0 t/s(4 cards)
- GLM-5.1 744B · 40B active · NVFP490.6 t/s(8 cards, speculative decoding)
- Qwen3.8 125B · 6B active · NVFP4101.6 t/s(speculative decoding, part in system RAM)
- Kimi K2.6112.5 t/s(8 cards, speculative decoding)
- GLM-5.3 320B · 18B active · EXL3-K3179.5 t/s(2 cards, speculative decoding)
- DeepSeek V4 Flash 284B · 13B active · FP4+FP8106.1 t/s(2 cards)
- Mimo 2.6 309B · 15B active246.0 t/s(2 cards, speculative decoding)
- Gemma 4 31B · NVFP451.0 t/s
- Muse 30B · BF1657.0 t/s(speculative decoding)
- Qwen3.8 27B · FP8114.8 t/s(speculative decoding, 4 requests at once)
- Qwen3.8 27B · 8-bit56.0 t/s(3 requests at once)
- Qwen3.8 125B · 6B active · NVFP4110.0 t/s(speculative decoding, part in system RAM)
Frequently asked
How fast is vLLM on the NVIDIA RTX Pro 6000 Blackwell?
On the NVIDIA RTX Pro 6000 Blackwell, vLLM runs Gemma 4 31B at 4-bit at a median 51.0 tokens per second for generation, from 1 plain run (16K context).