benchmarks / qwen3-6-35b-a3b--dgx-spark-x1--vllm--nvfp4--p2048-o256-c1--2026-07-06
Qwen3.6 35B-A3B on NVIDIA DGX Spark
compare76.3tok/s
323msttft
6.3kprefill
vLLM 0.24.0NVFP4Jul 6, 2026
Run configuration
nodes
1
batch size
—
concurrency
1
context
10240
prompt / output
2041 / 256
weights nvidia/Qwen3.6-35B-A3B-NVFP4 ↗
Metrics
decode tok/s
76.3
prefill tok/s
6.3k
ttft
323ms
total tok/s
69.9
inter-token
13ms
e2e latency
3.66s
req/s
0.273
peak memory
69.6 GB
avg power
30.4 W
energy/req
111.4 J
Reproduce
PATH="$HOME/venvs/vllm/bin:$PATH" VLLM_USE_DEEP_GEMM=0 TORCHINDUCTOR_COMPILE_THREADS=2 MAX_JOBS=4 vllm serve ~/models/hf/Qwen3.6-35B-A3B-NVFP4 --served-model-name nvidia/Qwen3.6-35B-A3B-NVFP4 --max-model-len 10240 --gpu-memory-utilization 0.5 --max-num-seqs 8 --limit-mm-per-prompt '{"image": 0, "video": 0}' --port 8000
python3 bench/harness.py bench/configs/qwen3-6-35b-a3b-nvfp4--vllm.json --rows-out data/benchmarks/qwen3-6-35b-a3b.jsonNVIDIA's official NVFP4 quant; tests native Blackwell FP4 tensor-core path (cutlass/flashinfer) on GB10.