AWQ (Activation-aware Weight Quantization)

AWQ, Activation-aware Weight Quantization from MIT HAN Lab, compresses large language and multimodal models to 3 or 4-bit weights by protecting the small fraction of salient weight channels identified from activation statistics, preserving more quality than naive quantization at the same bit-width. It pairs with the TinyChat runtime for efficient 4-bit inference, including vision-language models. Released under the MIT license.