Releasing MACaQ: one quantizer across modalities
Introducing MACaQ for text, images, audio, video and 3D, with an Apache 2.0 implementation, a paper and an interactive guide.
We are releasing MACaQ, short for Model Agnostic Calibrated Quantization: one quantization recipe for low-precision weights and activations, with examples spanning text, images, audio, video and 3D reconstruction.
The MACaQ implementation is released under the Apache 2.0 license.
One common building block
Different model architectures share a building block: the linear layer. MACaQ keeps balancing, rotation and quantization decisions local to that layer, avoiding architecture-specific changes to neighboring layers. Calibration and refinement prepare the weights offline; lightweight input transforms are applied during inference.
Explore the interactive blog
The MACaQ interactive blog brings the method to life through visual and audio comparisons, an explanation of the layer-local design, measured model-size and quality comparisons, and integration workflows for llama.cpp and vLLM. Image and video workflows also use Diffusers.
Read the paper on HF Papers for the full method and experiments, or open the interactive blog to explore the examples.
