Prefill and token generation speed
DGX Spark 128GB with Qwen3.6 35B-A3B (MoE): prefill and generation speed
Last updated
273 GB/s
Active 3B / Total 35B
61.2 ~ 111.1 tok/sDecode speed(baseline: No matching baseline)
Model fully loaded (24.32GB / 116GB)
First token (TTFT)
1.5 to 3.1 secondsContext 4,096 input tokens included
Prefill speed
1,341 ~ 2,661 tok/s1.5 ~ 3.1 s (Estimation based on public benchmarks)
Total time (E2E)
2.6 to 5.0 seconds117 output tokens included
Press Play to process the input and generate the answer.
Ready
Expand settings SettingsNVFP4 · 4K · NVFP4 + MTPCollapse
QuantizationNeed 24.32GB
vLLM · The Spark recipe uses Marlin for MoE and Triton for MTP. Selecting NVFP4 does not enable MTP automatically.
Input context4K tokens
NVFP4 + MTP
Baseline
Sample question:
117 output tokensCheck configuration
DGX Spark 128GB · Memory 128GB
Qwen3.6 35B-A3B (MoE) NVFP4 configuration. Check that the retailer lists the same memory option.
Already own it? Open the setup guideWould another device feel faster?
Compare the same model including prefill, or just token generation.
What changes with a longer document?Expand input length, waiting time, speed and memory
Detailed analysis and all 60 hardware configurations
Popular configurations
Speed simulations by device and model
Mac mini M4 Pro 64GB × Qwen3.6 35B-A3B (MoE)
19.6 ~ 46.1 tok/s · 2.7 ~ 6.3 s
Studio M4 Max 128GB × Qwen3.6 35B-A3B (MoE)
39.2 ~ 81.9 tok/s · 1.5 ~ 3.1 s
2× RTX 3090 48GB (NVLink) × Qwen3.6 35B-A3B (MoE)
79.2 ~ 248.6 tok/s · 0.63 to 1.4 seconds
Gemma 4 12B on Mac mini M4 24GB: chat speed
11.6 ~ 12.9 tok/s · 12.8 ~ 31.1 s