---
title: "Transformer compresses 1GB of Wikipedia to 105MB losslessly"
url: https://www.parallelquant.com/posts/transformer-compresses-1gb-of-wikipedia-to-105mb-losslessly-80e4ab
source_name: "Tom's Hardware"
source_url: https://www.tomshardware.com/tech-industry/artificial-intelligence/one-gigabyte-of-wikipedia-losslessly-compressed-to-105-39mb-with-self-learning-ai-transformer-model-tufazip-needs-two-nvidia-b200-gpus-and-about-a-week-to-compress-or-decompress
published: 2026-11-21T11:30:00.000Z
topics: ["research"]
publisher: "Parallel Quant"
---

# Transformer compresses 1GB of Wikipedia to 105MB losslessly

*2026-11-21 · Source: [Tom's Hardware](https://www.tomshardware.com/tech-industry/artificial-intelligence/one-gigabyte-of-wikipedia-losslessly-compressed-to-105-39mb-with-self-learning-ai-transformer-model-tufazip-needs-two-nvidia-b200-gpus-and-about-a-week-to-compress-or-decompress)*

Tufa Labs' tufazip compressed 1GB of Wikipedia to 105.39MB losslessly using a 163-million-parameter Transformer as its only predictor. It ran on two Nvidia B200 GPUs.

**Why it matters:** Compression and prediction are mathematically the same problem, so lossless Wikipedia benchmarks are a clean, hard-to-game measure of how well small models capture language. The result shows a compact neural predictor beating traditional hand-tuned compressors, though the GPU requirement keeps it a research milestone rather than a practical tool for now.

**Topics:** research

---
Read the original: https://www.tomshardware.com/tech-industry/artificial-intelligence/one-gigabyte-of-wikipedia-losslessly-compressed-to-105-39mb-with-self-learning-ai-transformer-model-tufazip-needs-two-nvidia-b200-gpus-and-about-a-week-to-compress-or-decompress
Canonical: https://www.parallelquant.com/posts/transformer-compresses-1gb-of-wikipedia-to-105mb-losslessly-80e4ab
Published by Parallel Quant — https://www.parallelquant.com
