LensVLM-9B by Apple

LensVLM-9B utilizes a post-training recipe and inference framework to maintain high accuracy in vision-language models by selectively expanding compressed image regions. Built on Qwen3.5-9B-Base, the model achieves performance comparable to full-text processing at 4.3x compression and outperforms existing baselines up to 10.1x across seven text QA benchmarks.