Researchers developed a quantization framework for deploying large vision-language models on memory-constrained mobile hardware, combining a self-generated training data pipeline that needs no access to the original training infrastructure with a new 2.7-bit parameter format optimized for ARM processors. The team compressed Meta’s Llama 3.2 11B Vision model to 3.7GB while maintaining strong performance on visual question-answering benchmarks. The approach targets efficient on-device inference for VLMs without requiring the original model developer’s training pipeline.
