LLM Inference Benchmark Explorer
Kimi K3 on DGX B300 — MXFP4, vLLM, TP8, speculative decoding inference benchmark
Open this configuration in the LLM Inference Benchmark Explorer
| Model | Kimi K3 |
|---|---|
| Parameters | 2.78T |
| Intelligence Index | 43.6 |
| Agentic Index | 50 |
| Device | DGX B300 |
| Quantization | MXFP4 |
| Max C | 32 |
| Chat Capacity | 128 |
| Agentic Capacity | 48 |
| TP | 8 |
| DP | — |
| PP | — |
| Engine | vLLM |
| Speculative Decoding | Yes |
| C | TTFT (ms) | TPS (tok/s) | Status |
|---|---|---|---|
| 1 | 71 | 187.74 | PASS |
| 2 | 173 | 129.75 | PASS |
| 4 | 191 | 99.43 | PASS |
| 8 | 263 | 61.04 | PASS |
| 16 | 344 | 43.60 | PASS |
| 32 | 565 | 23.14 | PASS |
| 64 | 1031 | 11.21 | FAIL |
In OpenZeka's measurement, Kimi K3 (2.78T parameters), served in MXFP4 format with vLLM and speculative decoding on DGX B300 (TP=8), reached a generation speed of 187.7 tok/s per request and a time to first token (TTFT) of 71 ms with a single request.
Considering the speed targets and the available KV cache capacity, the estimated capacity is 128 users for chat use and 48 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: DSpark speculative. Draft model: Inferact/Kimi-K3-DSpark. 1.86x speedup at C1.
Intelligence Index and Agentic Index values are published by Artificial Analysis and are reproduced here with attribution.
In OpenZeka's measurement, Kimi K3 (2.78T parameters), served in MXFP4 format with vLLM and speculative decoding on DGX B300 (TP=8), reached a generation speed of 187.7 tok/s per request and a time to first token (TTFT) of 71 ms with a single request.
Considering the speed targets and the available KV cache capacity, the estimated capacity is 128 users for chat use and 48 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.