Function expected_vision_tokens_bounded 

pub fn expected_vision_tokens_bounded(
    w: u32,
    h: u32,
    patch: u32,
    merge: u32,
    max_pixels: u64,
) -> u32

2026-09-26: Merged vision tokens for w × h under a declared area bound.

A serve with a bound downscales a larger image and answers (preprocess_image_with_max_pixels), so the prediction takes the bound as input. max_pixels == 0 means no bound was declared and returns expected_vision_tokens, on the premise that the checkpoint’s own bound (16777216 px in the Qwen3.6 preprocessor configs) is above every fixture. It is not served_size(.., None), which applies the 1280 px fallback clamp.