ShapeLearn is a quantization technology for large language models developed by ByteShape. Instead of applying a single bit width across the whole model, it uses a hybrid quantization approach that learns the most suitable data format for each tensor.
It aims to preserve model quality at very low bit widths, in the range of two to four bits. Models compressed with ShapeLearn are distributed in the GGUF format for local inference environments such as llama.cpp.