Neural model improves Wikipedia compression to 105MB
A self-learning transformer built by Tommy He compressed one gigabyte of Wikipedia content into 105.39 MB without losing data.

Key takeaways
- The compression was lossless, so no data was lost.
- The AI model served as the only predictor.
- A smaller tokeniser helped more than a larger one.
- The process required two Nvidia B200 GPUs and lengthy compute.
The compression result
Tom's Hardware reports that a self-learning AI model has advanced Wikipedia compression. Researcher Tommy He used a 163 million parameter Transformer to compress one gigabyte of Wikipedia content into 105.39 MB. The team described a method called tufazip, which produced a lossless file, and the work was shared on 19 November 2026.
How the model worked

The system did not act like a chatbot. It was a self-contained model that could learn as it went. A starter model was trained offline, then the full model began online learning. The unpacking process repeated the same steps, and predictor updates had to be deterministic.
What it means
The researchers found that a smaller vocabulary helped. A tokeniser with four thousand words performed better than larger options. They also found that a bigger model was not always better, as a two hundred sixty seven million parameter model scored slightly lower than a one hundred eighty eight million parameter model.
Limits and next steps
The work is described as research rather than a finished tool. Running it would require two Nvidia B200 GPUs and days of compute. The first online version came in at 105.98 MB, while the later version placed 13th in the benchmark.
Sources
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



