LLM Inference Benchmark Explorer
NVIDIA-Nemotron-3.5-Super-VL-120B-A12B on DGX B300 — BF16, vLLM, TP2, speculative decoding k=3 inference benchmark
Open this configuration in the LLM Inference Benchmark Explorer
| Model | NVIDIA-Nemotron-3.5-Super-VL-120B-A12B |
|---|---|
| Parameters | 120B |
| Intelligence Index | 26.5 |
| Agentic Index | — |
| Device | DGX B300 |
| Quantization | BF16 |
| Max C | 128 |
| Chat Capacity | 512 |
| Agentic Capacity | 192 |
| TP | 2 |
| DP | — |
| PP | — |
| Engine | vLLM |
| Speculative Decoding | Yes |
| C | TTFT (ms) | TPS (tok/s) | Status |
|---|---|---|---|
| 1 | 56 | 350.08 | PASS |
| 2 | 81 | 266.54 | PASS |
| 4 | 94 | 198.12 | PASS |
| 8 | 115 | 144.74 | PASS |
| 16 | 169 | 102.32 | PASS |
| 32 | 240 | 73.11 | PASS |
| 64 | 325 | 44.83 | PASS |
| 128 | 359 | 29.08 | PASS |
| 256 | 4014 | 17.42 | FAIL |
In OpenZeka's measurement, NVIDIA-Nemotron-3.5-Super-VL-120B-A12B (120B parameters), served in BF16 format with vLLM and speculative decoding on DGX B300 (TP=2), reached a generation speed of 350.1 tok/s per request and a time to first token (TTFT) of 56 ms with a single request.
Considering the speed targets and the available KV cache capacity, the estimated capacity is 512 users for chat use and 192 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 MTP k=3.
Intelligence Index and Agentic Index values are published by Artificial Analysis and are reproduced here with attribution.
In OpenZeka's measurement, NVIDIA-Nemotron-3.5-Super-VL-120B-A12B (120B parameters), served in BF16 format with vLLM and speculative decoding on DGX B300 (TP=2), reached a generation speed of 350.1 tok/s per request and a time to first token (TTFT) of 56 ms with a single request.
Considering the speed targets and the available KV cache capacity, the estimated capacity is 512 users for chat use and 192 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.