Qwen3.8-Flash-Next-NVFP4 MiaAI vLLM two-Spark baseline
2× DGX Spark · vLLM ·
Tested configuration
- Hardware
- 2x NVIDIA DGX Spark (128GB each)
- Serving runtime
- vLLM
- Runtime version or build
- vLLM 0.1.dev20073+g8e685d198, MTP speculative decoding with 3 draft tokens · vllm/vllm-openai:qwen38-flash-next
- Weight quantization or precision
- NVFP4
- Benchmark date
- Recipe credit
- MiaAI-Lab Qwen dual-Spark recipeBy MiaAI-Lab
169fbad266f2791335a3102f0d3d625e7c295563
Decode sweep
~256 input / 512 output tokens per request. Output throughput is aggregate across concurrent requests. TTFT means time to first token.
| Concurrency | Output tok/s | TTFT p95 (s) |
|---|---|---|
| 1 | 48.93 | 0.43 |
| 2 | 83.20 | 0.87 |
| 4 | 134.62 | 1.22 |
| 8 | 206.78 | 1.18 |
| 12 | 209.12 | 19.13 |
| 16 | 208.17 | 20.96 |
Prefill and context
- Configured request window
- 1,000,000 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/s | TTFT p95 (s) |
|---|---|---|
| 4,096 input tokens | 1,312.80 | 1.88 |
| 32,768 input tokens | 2,399.48 | 12.50 |
| 131,072 input tokens | 2,380.23 | 53.85 |
| 245,760 input tokensPasskey prompt | 2,069.12 | 117.30 |
Sustained load
20-minute soak at concurrency 8 · ~1024 input / 512 output tokens per request.
- Aggregate throughput
- 197.71 output tok/s
- TTFT p95
- 1.19 s
- Completed requests
- 469
- Request errors
- 0
Serving regression checks 13/13 passed
Tool-use evaluation
Unavailable · no published tool-eval result for this configuration.