Speed measurement details
The two figures on the landing page use different inputs and measurement boundaries. They are selected examples, not universal document processing times.
What the 1.18 seconds covers
One warm profile of a slide with 12 regions, recorded on September 11, 2026 on an M3 Pro with 18GB memory. The 1,177.58 ms excludes model download, initial loading, and final app rendering. It is one run, not a five-run average.
PP-DocLayoutV3 FP32 + MLX Q6/K16. The roughly 10 ms difference between the stage sum and total is counted as other time. Power and background-load conditions were not recorded.
What the 3.24× compares
Decoding 59 output tokens from a completed multimodal prefix cache to EOS. The 3.2382 ratio is rounded from the two-process central values: MLX BF16 AR at 877.12 ms and MLX Q6 K16 self-spec at 270.87 ms. Each process ran one warmup and five measurements per method.
Excludes layout, image loading and resizing, vision encoding, multimodal prefill, and tokenizer decoding. These diagnostic runs used battery power with Chrome active; clean runs on AC power are still needed. This is separate from the 1.18-second profile.