What we know
Tufa Labs has developed tufazip, a compression tool that uses a 163-million-parameter Transformer model running on two Nvidia B200 GPUs to losslessly compress 1GB of Wikipedia data down to 105.39MB. The reported compression ratio significantly exceeds what common algorithms achieve on similar datasets. Tufazip compresses 1GB of Wikipedia data to 105.39MB losslessly. It uses a 163M-parameter Transformer model as the only predictor. The compression process runs on two Nvidia B200 GPUs.
Why it matters
Reducing file sizes without losing any information is critical for efficient storage and transmission. Traditional compression tools like gzip or zstd depend on statistical models and heuristics, while AI-based methods are an emerging area of research. Tufa Labs’ use of a self-learning Transformer model alone for compression suggests a novel direction that could improve compression ratios for complex, large-scale datasets such as Wikipedia.
What is still unknown
No independent tests have confirmed tufazip’s compression accuracy or efficiency. Details about the range of supported languages, compatibility with other data types, and plans for public release remain undisclosed. The computational cost and speed of compression and decompression beyond the Nvidia B200 GPU setup are also unclear.

