1× RTX 5090
252tok/sEveryday
1 request, prose prompt, no speculative decoding Source
Intelligence18with thinkingno thinking 15
Local AI · Recipe · 1× DGX Spark
Qwen3.6 27B NVFP4 with vLLM on 1× DGX Spark
Qwen3.6 27B NVFP4 with vLLM on 1× DGX Spark: 45.6 tok/s according to github.com (dataset as of Nov 3, 2026).
by MiaAI-Lab
Intelligence (original model) · Artificial Analysis21with thinkingno thinking 20
Engine
- Engine
- vLLM nightly-aarch64 (ungepinnt)
- Quantization
- NVFP4 weights + bf16 KV
- Model family
- Qwen3.6
- Context
- 262,144
- Parameters
- ~27B dense
- Creator
- MiaAI-Lab
- GitHub stars
- 22
- Repo updated
- Sep 16, 2026
Measurements
Every sourced value of this recipe, each with its condition and source. Bars relative to the largest value in the group.
Decode · 1 request
Prefill by context
Time to first token
Shorter is better.
What you need
- Hardware
- 1 × NVIDIA DGX Spark (GB10)
- Weights
- Main weightsnvidia/Qwen3.6-27B-NVFP4 Apache-2.0
- Draft modelz-lab/Qwen3.6-27B-DFlash MIT
- Engine
- vLLM nightly-aarch64 (ungepinnt)
- Context
- 262,144 tokens
Notes
What matters before you rebuild it.
- Unpinned nightly
Sources
Related recipes
2× RTX 5090
249tok/sEveryday
1 request, prose prompt, no speculative decoding Source
Intelligence18with thinkingno thinking 15
- License unclear
2× RTX 3090
179tok/sEveryday
1 request, prose prompt, no speculative decoding Source
Intelligence18with thinkingno thinking 15
- License unclear
2× RTX 5090
168tok/sEveryday
1 request, prose prompt, with MTP Source
Intelligence21with thinkingno thinking 20
2× RTX 5090
153tok/sEveryday
1 request, prose prompt, with MTP (per recipe) Source
Intelligence21with thinkingno thinking 20
1× RTX 5090
128tok/sEveryday
1 request, prose prompt, with MTP (per recipe) Source
Intelligence21with thinkingno thinking 20
Local AI in your company?
In a workshop we work out which models and which hardware fit your tasks, and build the first agent on your infrastructure.