x0ne Engine · benchmarks
x0ne Core Confidential Inference Speed
One NVIDIA H100 80 GB in confidential mode, in an Intel TDX VM on Google Cloud (a3-highgpu-1g). x0ne Core, our GPU engine, 4 November 2026. One request at a time, prompts of random token ids, greedy decoding, 256 tokens generated, n = 5 per cell, Qwen3 weights in bf16.
| Prompt of 128 tokens | Prompt of 16,384 tokens | |||
|---|---|---|---|---|
| Model | Generation, tokens/s | First token, s | Generation, tokens/s | First token, s |
| Qwen3 0.6B | 730 | 0.0035 | 385 | 3.15 |
| Qwen3 8B | 138 | 0.0089 | 117 | 7.94 |
| Qwen3 14B | 81 | 0.0147 | 66 | 11.08 |
| Qwen3 32B | 38 | 0.0317 | 35 | 27.39 |
Through x0ne Engine’s relay, which verifies every receipt: Qwen3 8B at 137 tokens/s (prompt of 58 tokens, 1,024 generated); 36 receipts verified out of 36, warm-up included; test signing keys.