GLM-5.3 Flash EXL3/TR3 4 bpw refreshed 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.dev20051+g487ecf187
- Weight quantization or precision
- EXL3/TR3 4 bpw
- Benchmark date
- Recipe credit
- MiaAI-Lab refreshed GLM recipeBy MiaAI-Lab
a35eaab128233d215b2fc9ac261eddbad23c946d
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 | 13.96 | 1.27 |
| 2 | 22.68 | 15.26 |
| 4 | 30.68 | 10.48 |
| 8 | 25.60 | 125.72 |
| 12 | 27.61 | 173.89 |
| 16 | 26.88 | 267.50 |
Prefill and context
- Configured request window
- 850,000 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 | 420.55 | 5.27 |
| 32,768 input tokens | 781.30 | 37.52 |
| 131,072 input tokens | 873.52 | 147.68 |
| 245,760 input tokensPasskey prompt | 843.89 | 287.63 |
Sustained load
20-minute soak at concurrency 8 · ~1024 input / 512 output tokens per request.
- Aggregate throughput
- 21.88 output tok/s
- TTFT p95
- 181.03 s
- Completed requests
- 56
- Request errors
- 0
Serving regression checks 13/13 passed
Tool-use evaluation
Unavailable · no published tool-eval result for this configuration.