How to read these results

A useful number needs its settings, its source and its limits.

Start with the exact configuration

Each report identifies the hardware, runtime, model and test where recorded. A “cell” is one combination of those settings. Different cells in the same campaign may use different models, memory offload or reasoning budgets. A faster cell is not automatically a better assistant.

What the timing numbers mean

• Tokens per second (tok/s): how quickly the runtime generated tokens during the measured call. Token definitions and timing boundaries can differ between engines.
• Duration: elapsed time for the recorded request. A short answer can finish sooner despite lower generation speed.
• p50: the median observed value. Half the observations are at or below it. p95: the 95th percentile; with small samples this is only a rough description.
• Cold / warm: the harness labels the first call in a sequence “cold” and later calls “warm”. Check the source before treating this as a verified disk reload or cache reset.
• Missing data: a dash or “not recorded” means no supported measurement. It does not mean zero.

Speed and correctness are separate

Some older reports label nonempty output as a “usable answer”. That only confirms the model returned text. “Marker coverage” checks expected words or phrases; it is not a code test or a correctness score. Formatting checks are also separate from whether an answer is right. Read these as limited automated checks, then inspect the outputs.

What is comparable?

Use the same prompt suite, model revision, quantization, context size, output limit and runtime settings. Record background load, caching and memory limits where possible. Tests made on different dates or with different quantizations cannot isolate the cause of a speed difference. In particular, the Pavilion and Predator A3B reports use different IQ2 variants.

What the catalogue contains

The catalogue contains 145 records: 26 selected reports or campaign groups, 16 archive-only records and 103 historical documents. A document can summarize multiple runs; a media group can contain many attempts. These are not 145 independent, directly comparable tests. Model and hardware directories are indexes of mentions, not performance rankings.

Evidence and reproducibility

Reports link to retained JSONL records, summaries and metadata where available. Some historical records lack logs, source revisions or complete raw data; those gaps are stated at the link. Downloadable publication copies remove private paths and network addresses. See data and attribution and the complete catalogue.

Images, video and music

The October 2026 media campaign used a request pipeline ending in ComfyUI on a GTX 1060 6 GB. The dataset preserves prompts, seeds, workflow information, timings, returned media and failed or cancelled attempts where captured.

Native graph duration measures work inside the generation graph. Factory trace duration covers the worker trace and adapter overhead. Delivery time can also include transfer and queueing. These clocks are reported separately and are not interchangeable.

Repeated seeds, conditioning caches and changed workflows can affect the timings. A retained or decoded file proves delivery, not good image quality, useful motion or musical quality. Q8 describes weight storage; it does not prove 8-bit compute or dedicated-VRAM residency under Windows.

Inspect the media gallery, the full dataset and the plain-language findings. Original private operator logs and credentials are excluded.