2026-04-13 Pavilion Windows GPU Qwen3.5 0.8B Long Context
TL;DR
This report measures how long the listed long-input scenarios took on the Pavilion GPU. Larger context settings can increase prompt processing and memory costs substantially. Completed timing rows do not establish that the model used or recalled the entire context correctly.
Generated 2026-04-13 14:55:26 . Source run: C:\CodexProjects\MyServers\instances\pavilion-windows-laptop\telemetry\generated\long-context\qwen35-0.8b\2026-04-13_14-01-26\gpu
| Scenario | Context Tokens | Prompt Tokens | Prefill Seconds | Prefill tok/s | Generation Tokens | Generation Seconds | Generation tok/s | Wall Seconds |
|---|---|---|---|---|---|---|---|---|
| gpu-016k | 16,384 | 15,565 | 27.266 | 570.857 | 128 | 4.373 | 29.271 | 35.524 |
| gpu-032k | 32,768 | 31,633 | 72.232 | 437.936 | 128 | 4.77 | 26.834 | 81.047 |
| gpu-050k | 50,000 | 50,000 | 148.497 | 336.707 | 128 | 55.842 | 2.292 | 208.916 |
| gpu-080k | 80,000 | 80,000 | 326.531 | 245.0 | 128 | 109.823 | 1.166 | 442.612 |
| gpu-128k | 131,072 | 131,072 | 836.292 | 156.73 | 256 | 293.302 | 0.873 | 1138.987 |