tenz losslessly compresses GGUFs by up to 13%, saving 80-130GB per 1TB archive

Simpli-curious · reddit · 2026-11-16

A Reddit developer released tenz 0.25, a lossless compressor for GGUF/safetensors model files—decompressed files are byte-for-byte identical, so llama.cpp, Ollama and LM Studio load them as before. It's a single dependency-free C file, developed and benchmarked on a Poco F5 phone in Termux.

Results

How it works: general compressors fail on quantized weights because they look random; tenz parses the GGUF, unpacks each block into real quantized values and scales, and codes them with context models fitted to the file, using rANS from static tables so decoding needs no learning. An experimental OpenCL GPU path on the phone's Adreno 725 hit 150 MiB/s—still slower than the big CPU cores (250 MiB/s)—and had to work around a driver byte-store bug.

Caveats: pre-release files only decompress with the creating version; decoding is slower than zstd (7s per 1.2GB on phone); gains apply mainly to model files. Code and full tables on GitHub.

Original post →

More from coding & agent

coding & agent channel →