Downloads
- benchmarks.csv: one row per measurement (model, hardware, settings, tokens per second).
- benchmarks.json: the same results with full test conditions and machine details.
- models.json: file size, active bytes per token and conversation-memory cost for popular models, read from the model files. This powers our calculator.
Raw output
The unedited llama-bench output for every test session, plus the system details our script recorded:
bench-desktop-20260925-184928
- cpu-gemma-4-e4b.json
- cpu-gpt-oss-20b.json
- cpu-llama-2-7b.json
- cpu-qwen3.5-9b.json
- cuda-gemma-4-e4b.json
- cuda-gpt-oss-20b-cpumoe.json
- cuda-gpt-oss-20b-fit.json
- cuda-llama-2-7b.json
- cuda-qwen3.5-9b.json
- system.json
bench-k1-20260926-130154
- cpu-gemma-4-e4b.json
- cpu-gpt-oss-20b.json
- cpu-llama-2-7b.json
- cpu-qwen3.5-9b.json
- system.json
- vulkan-gemma-4-e4b.json
- vulkan-gpt-oss-20b.json
- vulkan-llama-2-7b.json
- vulkan-qwen3.5-9b.json
License and credit
Our data is licensed under Creative Commons Attribution 4.0. Please credit it as "Source: AIPCs (aipcs.store)" with a link to the page you used. It contains only our own measurements and calculations, no retailer data.