kindlingai/exlr8 is a low-bit model quantization format tuned for DGX Spark. According to the repository description, it focuses on reducing the bytes of expert weights read per token when serving models on unified memory hardware. There is no public latest release.
MoE is a structure that processes by activating only some experts depending on the input. EXLR8 took the cost of reading the weights of active experts, rather than compression ratio, as its design criterion. Building on the trellis code of the EXL3/QTIP family, it rounds codebook values to FP8 E4M3 during Viterbi search so that the reconstructed weights match FP8 values. It also stores K2/K3/K4 encoding widths per expert and selects the width through a manifest at load time. Instead of rotating each expert separately, it uses a single layer-level rotation so that the 8 active experts share the token activation rotation. The README introduces the format specification, a pure PyTorch reference loader, and a Python fileset writer.
The 2.3~33x advantage presented by the repository is a claim based on its own benchmark table. There are no tests and no license notice at the root, so the evidence for confirming quality and usage conditions is limited. The design can therefore be referenced, but there is insufficient verification basis to accept the presented performance as a general advantage.