LLM Inference Benchmark Explorer
Qwen3-VL-30B-A3B-Instruct on Thor — FP8, vLLM, TP1, speculative decoding inference benchmark
Open this configuration in the LLM Inference Benchmark Explorer
| Model | Qwen3-VL-30B-A3B-Instruct |
|---|---|
| Parameters | 30B |
| Intelligence Index | 9.5 |
| Agentic Index | — |
| Device | Thor |
| Quantization | FP8 |
| Max C | 8 |
| Chat Capacity | 32 |
| Agentic Capacity | 12 |
| TP | — |
| DP | — |
| PP | — |
| Engine | vLLM |
| Speculative Decoding | Yes |
| C | TTFT (ms) | TPS (tok/s) | Status |
|---|---|---|---|
| 1 | 75 | 55.98 | PASS |
| 2 | 109 | 41.07 | PASS |
| 4 | 149 | 27.47 | PASS |
| 8 | 189 | 20.83 | PASS |
| 16 | 225 | 15.63 | FAIL |
| 32 | 300 | 12.32 | FAIL |
In OpenZeka's measurement, Qwen3-VL-30B-A3B-Instruct (30B parameters), served in FP8 format with vLLM and speculative decoding on Thor, reached a generation speed of 56 tok/s per request and a time to first token (TTFT) of 75 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: Speculative decoding used; mechanism and depth not recorded. Model: Qwen/Qwen3-VL-30B-A3B-Instruct-FP8. 256K context
Intelligence Index and Agentic Index values are published by Artificial Analysis and are reproduced here with attribution.
In OpenZeka's measurement, Qwen3-VL-30B-A3B-Instruct (30B parameters), served in FP8 format with vLLM and speculative decoding on Thor, reached a generation speed of 56 tok/s per request and a time to first token (TTFT) of 75 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.