Tufa Labs' tufazip compresses 1GB of Wikipedia text to 105.39MB using only a neural network
Researcher Tommy He's tufazip tool, built at Tufa Labs, losslessly compressed a 1GB Wikipedia text sample down to 105.39MB using a 163-million-parameter Transformer model as its sole predictor, with no hand-built compression models involved. The package includes a small 4.33MB quantized starter model that a 12-layer Transformer expands and continues training on during decompression, a process that would take two Nvidia B200 GPUs roughly a week.
GoKawiil's interpretation of the reporting above, not reported fact.
The result tops the Large Text Compression Benchmark among entries relying purely on a neural network predictor, suggesting pure-transformer approaches can now rival systems that blend neural networks with traditional hand-built compression models. The heavy GPU time needed for compression and decompression indicates this remains a research demonstration rather than a practical tool, since speed and compute cost would need to improve dramatically for real-world use.
- tufazip compressed 1GB of Wikipedia text to 105.39MB losslessly using only a Transformer model.
- It ranks as the best neural-network-only entry on the Large Text Compression Benchmark.
- Compression and decompression each take about a week on two Nvidia B200 GPUs, limiting practical use.
Source: tomshardware.com — Shane Downing, 2026-11-21
Published there as: “One gigabyte of Wikipedia losslessly compressed to 105.39MB with self-learning AI transformer model”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.