LLM Inference Benchmark Explorer
Qwen3.5-397B-A17B on 4× DGX Spark — INT4, vLLM, TP4 inference benchmark
Open this configuration in the LLM Inference Benchmark Explorer
| Model | Qwen3.5-397B-A17B |
|---|---|
| Parameters | 397B |
| Intelligence Index | 21.4 |
| Agentic Index | — |
| Device | 4× DGX Spark |
| Quantization | INT4 |
| Max C | 8+ |
| Chat Capacity | 32+ |
| Agentic Capacity | 12+ |
| TP | 4 |
| DP | — |
| PP | — |
| Engine | vLLM |
| Speculative Decoding | — |
| C | TTFT (ms) | TPS (tok/s) | Status |
|---|---|---|---|
| 1 | 325 | 41.88 | PASS |
| 2 | 393 | 33.40 | PASS |
| 4 | 541 | 27.16 | PASS |
| 8 | 622 | 20.57 | PASS |
In this measurement, Qwen3.5-397B-A17B (397B parameters), served in INT4 format with vLLM on 4× DGX Spark (TP=4), reached a generation speed of 41.9 tok/s per request and a time to first token (TTFT) of 325 ms with a single request.
Considering the speed targets and the available KV cache capacity, the estimated capacity is at least 32 users for chat use and at least 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.
“At least” means the configuration met the speed targets even at the highest load tested, so the real capacity may be higher.
Notes: Switch topology, TP=4. 2.24x faster than PP=3 at C1.
Intelligence Index and Agentic Index values are published by Artificial Analysis and are reproduced here with attribution.
In this measurement, Qwen3.5-397B-A17B (397B parameters), served in INT4 format with vLLM on 4× DGX Spark (TP=4), reached a generation speed of 41.9 tok/s per request and a time to first token (TTFT) of 325 ms with a single request.
Considering the speed targets and the available KV cache capacity, the estimated capacity is at least 32 users for chat use and at least 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.
“At least” means the configuration met the speed targets even at the highest load tested, so the real capacity may be higher.