Spark Bench

Qwen3.6-35B-A3B NVFP4 (Unsloth): vLLM GB10 + MiaAI-Lab patch, MTP 2, one Spark, 262K, thinking off

1× DGX Spark · vLLM ·

Tested configuration

Hardware
1x NVIDIA DGX Spark (128GB)
Serving runtime
vLLM
Runtime version or build
vLLM 0.26.1.dev0+g568afb3a1 with FlashInfer 0.6.14, built locally from timothystewart6/vllm-gb10 @ 94352d3 with its pinned versions
Weight quantization or precision
NVFP4
Benchmark date
Recipe credit
MiaAI-Lab Unsloth Qwen3.6-35B NVFP4 DGX Spark recipeBy MiaAI-Lab79c9e6f359f6101cdacd0dfd6fe9861ae2493a4d

Decode sweep

~256 input / 512 output tokens per request. Output throughput is aggregate across concurrent requests. TTFT means time to first token.

ConcurrencyOutput tok/sTTFT p95 (s)
177.460.20
2114.901.33
4190.660.35
8306.341.24
12343.870.67
16429.210.95

Prefill and context

Configured request window
262,144 tokens
Longest tested prompt
131,072 target input tokens
Validated retrieval
UnavailablePasskey not run

Input throughput uses full request wall time. The passkey measures retrieval from one long prompt.

Workload (c1)Input tok/sTTFT p95 (s)
4,096 input tokens2,710.940.65
32,768 input tokens3,711.678.49
131,072 input tokens2,429.7753.44

Sustained load

20-minute soak at concurrency 8 · ~1024 input / 512 output tokens per request.

Aggregate throughput
286.66 output tok/s
TTFT p95
0.52 s
Completed requests
678
Request errors
0

Serving regression checks 12/12 passed

Tool-use evaluation

Complete · 100% completionSafety gate failed

Compare only against complete results with 100% completion under the same protocol ID.

Protocol ID
tool-eval-standard-v1-b60e2f0
Workload
69 mock-tool scenarios per trial · 3 trials
Mean score
88.3 / 100
Pass@3 · passed at least once
89.9%
Pass³ · passed every trial
76.8%

Read the tool-eval report and scenario findings