LLM Inference Benchmark Explorer
Qwen3.8-27B on 1× DGX Spark — NVFP4, SGLang, TP1, speculative decoding k=8 inference benchmark
Open this configuration in the LLM Inference Benchmark Explorer
| Model | Qwen3.8-27B |
|---|---|
| Parameters | 27B |
| Intelligence Index | 33.7 |
| Agentic Index | 45.8 |
| Device | 1× DGX Spark |
| Quantization | NVFP4 |
| Max C | 8 |
| Chat Capacity | 32 |
| Agentic Capacity | 12 |
| TP | — |
| DP | — |
| PP | — |
| Engine | SGLang |
| Speculative Decoding | Yes |
| C | TTFT (ms) | TPS (tok/s) | Status |
|---|---|---|---|
| 1 | 232 | 47.94 | PASS |
| 2 | 320 | 39.80 | PASS |
| 4 | 343 | 35.60 | PASS |
| 8 | 389 | 27.64 | PASS |
| 16 | 2357 | 18.62 | FAIL |
| 32 | 9388 | 9.47 | FAIL |
In OpenZeka's measurement, Qwen3.8-27B (27B parameters), served in NVFP4 format with SGLang and speculative decoding on 1× DGX Spark, reached a generation speed of 47.9 tok/s per request and a time to first token (TTFT) of 232 ms with a single request.
Considering the speed targets and the available KV cache capacity, the estimated capacity is 32 users for chat use and 12 for agentic use. The realistic capacity will likely fall between these two values. In scenarios dominated by coding, tool use, long workflows and multi-agent use, capacity approaches the agentic estimate; where shorter interactions, standard conversations and lighter tasks dominate, it approaches the chat estimate.
Notes: DFlash2 Ultimate, BF16-LMHead, k=8, mem-fraction 0.50, KV fp8_e4m3
Intelligence Index and Agentic Index values are published by Artificial Analysis and are reproduced here with attribution.
In OpenZeka's measurement, Qwen3.8-27B (27B parameters), served in NVFP4 format with SGLang and speculative decoding on 1× DGX Spark, reached a generation speed of 47.9 tok/s per request and a time to first token (TTFT) of 232 ms with a single request.
Considering the speed targets and the available KV cache capacity, the estimated capacity is 32 users for chat use and 12 for agentic use. The realistic capacity will likely fall between these two values. In scenarios dominated by coding, tool use, long workflows and multi-agent use, capacity approaches the agentic estimate; where shorter interactions, standard conversations and lighter tasks dominate, it approaches the chat estimate.