Serve images with VISION=1 (#23)
Verified on current main: 0.858 GiB tower from the safetensors headers, 15.88 vs 15.02 GiB model loading, identical 68,605-token KV pool, OCR round-trip reads the pixels with VISION=1 and 400s cleanly with VISION=0.
M
mhenrichsen committed
1d36852d5762b0ca2cc4101144cbdb3d2829636a
Committed by GitHub <noreply@github.com>
on 8/23/2026, 12:48:20 PM