One gigabyte of Wikipedia losslessly compressed to 105.39MB with self-learning AI transformer model
Reported by Tom's Hardware1 source ↓
Tufa Labs’ tufazip compressed 1GB of Wikipedia text to 105.39MB losslessly, using a Transformer as its only predictor. The result is a research benchmark entry, not a product.
Key points
- The compressed package includes a 4.33MB starter model needed to decompress the text
- Compressing the data took about 7.3 days on two Nvidia B200 GPUs; decompression took about 7.5 days
- The result ranked 13th on the benchmark as of its Nov. 19 update
Details
Tom’s Hardware reports that researcher Tommy He built tufazip for the Large Text Compression Benchmark. The 105.39MB total includes the compressed text, the unpacking program and its required files, including a quantized 4.33MB starter model; it does not include the full model learned during compression.
The system starts with a 34-million-parameter model trained offline, then learns while predicting text tokens. Decompression repeats the same deterministic updates as it rebuilds the model, allowing the text to be recovered without data loss. Tufa Labs says the entry uses a neural network as its only predictor, unlike higher-ranked entries that combine Transformers with hand-built models.
On the test machine, compression took 631,548 seconds and decompression 645,748 seconds, using two Nvidia B200 GPUs and two Intel Xeon Platinum 8570 CPUs. The earlier online-only version produced a 105.98MB file and ranked second when submitted in early September; the 105.39MB version ranked 13th in the benchmark’s Nov. 19 update.
Written from coverage as of
Images
- Tom's Hardware
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh