Function expected_vision_tokens 

pub fn expected_vision_tokens(w: u32, h: u32, patch: u32, merge: u32) -> u32

2026-09-26: Merged vision tokens for a w × h image at native size.

Both sides snap to patch × merge, then each patch × patch square is one patch and each merge × merge block of patches one token. At patch 16, merge 2 a 448×448 image is 784 patches and 196 tokens; 196 was measured against a live server on 2026-08-14.