Lossless Model Compression

Martín Eduardo Bravo Díaz
Undergraduate thesis, Department of Computer Science, University of Chile, 2026.

Advisors: Andrés Abeliuk and Gonzalo Navarro.

Download the thesis (PDF) Official DCC copy

Abstract

Foundation models, including large language models, image-generation models, and vision classifiers, have grown substantially in both capability and computational requirements. Reducing the cost of storing and transferring their weights has therefore become an important problem. Lossy techniques such as quantization and distillation can produce much smaller representations, but they modify the model and may affect its quality. Lossless compression addresses a different objective: reducing storage while reconstructing the original bits of every tensor exactly.

This thesis presents NeuralZip, a compressor for foundation-model weights inspired by compact data structures. The method exploits blockwise heterogeneity and local exponent redundancy. NeuralZip clusters tensors with similar exponent distributions, reuses Huffman codes within each cluster, and selects packed-exponent representations when their compression benefit justifies their measured encoding cost, while preserving bitwise-exact reconstruction.

The evaluation compares NeuralZip with ZipNN across language, vision, diffusion, and mixture-of-experts models, as well as training-state and federated-communication settings. In direct model storage, NeuralZip encodes faster in 26 of 34 model artifacts, ranging from near parity to 21.51×. Every floating-point archive reconstructs the original bits exactly.

Citation

@mastersthesis{bravo2026lossless,
  title  = {Lossless Model Compression},
  author = {Bravo Díaz, Martín Eduardo},
  school = {University of Chile},
  year   = {2026}
}