Braco compresses visual tokens 64x with almost no accuracy loss

Braco shrinks the standard 576 visual tokens to just 9 — a 64x reduction — while retaining 91.2–95.2% of the original model's accuracy. Prefill FLOPs drop by up to 86%, yielding roughly 30% end-to-end speedup.

• 576 tokens compressed to 9 — a 64x reduction
• Accuracy stays at 91.2–95.2%
• Prefill FLOPs cut by up to 86%
• About 30% end-to-end inference speedup