Compressing An 11B VLM To 2.7-bit Weights For Mobile CPUs

Combining a novel weight format for efficient decoding with quantization-aware training to reduce model size while retaining accuracy for multimodal AI.