
Tufa Labs tufazip compresses 1 GB of Wikipedia to 105.39 MB losslessly
Tufa Labs says tufazip losslessly compressed 1 GB of Wikipedia to 105.39 MB using a self-learning Transformer predictor, without hand-built models, but it needs two Nvidia B200 GPUs for about a week.
Tufa Labs says its tufazip system losslessly compressed 1 GB of Wikipedia into 105.39 MB. The method uses a 163-million-parameter Transformer as its only predictor and no hand-built compression model. It is the smallest result on the Large Text Compression Benchmark among entries that use a neural network as the sole predictor.
How tufazip works
The full package is 105.39 MB after optimization and includes a small, 4.33 MB pre-trained starter model stored at low precision, or quantized, to save space; that starter is needed to unpack the file. The full model is a 12-layer Transformer, while the starter stores just two already-trained stages and copies them back and forth to fill all 12 once the full model spins up in memory; each copy keeps learning on its own. This is a self-contained language model that learns on the fly, not a traditional chatbot or cloud AI, and the compression is lossless, so predictor updates must be fully deterministic. Training happens in two stages: a 34-million-parameter starter is trained offline before compression begins, then the full model learns online by predicting each token and training on it, with the unpacking side repeating the same steps and rebuilding the model alongside the text. The total size includes the compressed text, the unpacking program, and needed files including the starter, but not the full trained model. Tufa Labs says this sets it apart from all 12 entries that now rank above it on the benchmark, which mix a Transformer with hand-built models.
Performance and hardware needs
Researcher Tommy He built tufazip, and a research post dated November 19, 2026, introduced it. Its first, online-only version ranked second when submitted in early September; the 105.39 MB version was 13th as of the benchmark's November 19 update. The benchmark is text-only and the work is for research purposes, not a product. Compression and decompression would need two Nvidia B200 GPUs plus two Intel Xeon Platinum 8570 CPUs in the test machine, taking 631,548 seconds and 645,748 seconds, or about 7.3 and 7.5 days, not counting the starter's offline training. The first version, at 105.98 MB, took 647,595 and 630,136 seconds. GPU VRAM peaked at 43.37 GiB between them, and total memory use, GPUs included, peaked at roughly 68,525 MiB during compression and 73,748 MiB during decompression. For comparison, cmix-neif 2.0 reached 89.35 MB on an AMD Epyc 7B13 with four cores in 49.2 hours compressing and 49.3 hours decompressing; DNA-4 reached 105.55 MB on an Nvidia RTX 4090 in 19.8 hours including training and 2.6 hours decompressing; NNCP v3.2 reached 107.26 MB on an Nvidia RTX 3090 in 67.2 hours and 66.3 hours.
Experiments and trade-offs
The experiments show that a smaller vocabulary can help: a 4K-word tokenizer beats both 8K and 16K. Giving Wikipedia's XML markup its own dedicated tokens made compression worse. A 267-million-parameter model did slightly worse than a 188-million-parameter model on this data set, and with a smaller 100 MB version of the file, gains leveled off at about 77 million parameters. A run started from a bigger, six-layer offline model predicted better, with a 99.17 MB payload, but its chunkier weights made the total file larger than the online-only run. Widening the window of already-processed text used for extra training from 15 to 35 million tokens cut the payload only 1.21 percent at the cost of 40 percent more compute. Tufa Labs presents optimization as not always as simple as adding more resources, with trade-offs that can be more expensive than expected. He frames the six-layer trade-off as the next research question: how much predictive quality a smaller model can keep, making the vanishing point a balance between predictive quality and model size for this benchmark.
Sources: Tom's Hardware
SiTech — AI-powered web development
We build fast, modern websites and bring AI into real business workflows. Have a project or a question? We'd love to help.