Spark Bench

GLM-5.3 Flash NVFP4 vLLM two-Spark low-reasoning baseline

2× DGX Spark · vLLM ·

Tested configuration

Hardware
2x NVIDIA DGX Spark (128GB each)
Serving runtime
vLLM
Runtime version or build
vLLM in MiaAI-Lab image mia/glm53-flash-spark:mm-ray-v1
Weight quantization or precision
NVFP4
Benchmark date
Recipe credit
MiaAI-Lab GLM-5.3 Flash NVFP4 recipeBy MiaAI-Labaed98a13ca75140d2691cc5c651ea5817d9a3e44

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)
114.200.45
227.400.97
440.876.43
865.941.88
1263.0469.96
1666.1863.40

Prefill and context

Configured request window
262,144 tokens
Longest tested prompt
245,760 target input tokens
Validated retrieval
≈240K input tokens · passkey passed

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 tokens560.422.84
32,768 input tokens1,096.5728.40
131,072 input tokens1,252.93100.93
245,760 input tokensPasskey prompt1,288.72188.82

Sustained load

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

Aggregate throughput
60.90 output tok/s
TTFT p95
4.22 s
Completed requests
144
Request errors
0

Serving regression checks 13/13 passed

Tool-use evaluation

Unavailable · no published tool-eval result for this configuration.

Other configurations for GLM-5.3 Flash