Wan 2.2 gets post-training quantization to W4A4

View on Reddit

A post-training quantization technique reduces Wan 2.2 I2V model to W4A4 precision, combining weights and activations into a single model. The quantized model is available on HuggingFace, with code on GitHub, aiming to reduce memory usage while maintaining quality.