Tufazip v2, a lossless compression program developed by Tommy He of AI research company Tufa Labs, compressed 1,000,000,000 bytes (1 GB) of data extracted from English Wikipedia to 105,388,353 bytes (about 105.39 MB), including the files needed for decoding. Submitted on September 30, 2026, the method combines Transformer-based prediction with online learning and can restore the original data perfectly. According to a research report dated October 14, introducing a pretrained model reduced the compressed data itself by about 4.97 MB compared with the first version, submitted on September 7. However, the model needed for decoding grew larger, so the final reduction was only about 0.59 MB.
The method's distinguishing feature is that the decoder, not just the compressor, trains an AI model as it reads the data. This can improve the compression ratio, but decoding also demands enormous computation: restoring the 1 GB of data takes about 7.5 days.
How the decoder trains an AI model too
Tufazip splits text into tokens in a way that allows exact restoration of the original byte sequence, then uses a Transformer to predict the probability of the next token. The data is encoded on the basis of those predictions using "arithmetic coding." In arithmetic coding, each token is assigned an interval whose width corresponds to its predicted probability, and the interval matching the token that actually appears is selected in turn. Ideally, if the predicted probability of the token that appears doubles, the information needed to encode it falls by one bit. If the decoder can calculate the same probabilities, it can accurately recover the original tokens from the compressed bit stream.
He drew on NNCP, a lossless compression program that combines a Transformer with online learning. In tufazip as well, the compressor and decoder start from the same initial weights and use the tokens read so far to predict the next one. When the decoder reads a token the compressor encoded, both sides use the same data to update the model's weights and move on to the next prediction. If probability calculation and training are deterministic, meaning the same input and state always produce the same result, the decoder can reproduce the same training outcome without the updated model being sent.
Tufazip does not only learn from new data. It also uses "replay," repeatedly training on text it has already processed. Because the decoder has the same text, no additional training data needs to be stored, but the computational load rises with the number of training passes. In a comparison using a 163-million-parameter model and a 4K-vocabulary tokenizer, increasing the replay window from 15 million to 35 million tokens reduced the compressed data from 107.22 MB to 105.92 MB. That was a reduction of 1.30 MB (1.21%), while the estimated computation for the model, including online learning and replay, rose by about 40%.
In v2, a configuration called "Cycle2 (C2)" is used to store the pretrained initial model compactly. Two pretrained Transformer blocks are alternately duplicated to build a 12-layer model, and each layer is trained independently during operation. The finished model has 163 million parameters, but the initial weights handed to the decoder are held to 4,334,741 bytes through quantization. Even if pretraining improves prediction accuracy, overall size does not fall unless the compressed data shrinks by more than the storage cost of those weights.
Compressed data down about 4.97 MB, total size down about 0.59 MB
The Large Text Compression Benchmark, in which tufazip participated, is a lossless compression benchmark run by Matt Mahoney. Its target, "enwik9," is the first 1,000,000,000 bytes of an English Wikipedia XML dump dated March 3, 2006. Rankings are determined not only by the size of the compressed data but by the combined size of that data and the ZIP file containing the decompression program, any required dictionary, model weights and so on. In other words, simply moving decoding information into a separate file does not count as an improvement in compression.
According to the benchmark's official records, the results for the first version and v2, both submitted in September 2026, were as follows.
| Item | First version (v1) | v2 |
|---|---|---|
| Compressed data | 105,924,360 bytes | 100,958,324 bytes |
| Decoder ZIP | 54,709 bytes | 4,430,029 bytes |
| Total | 105,979,069 bytes | 105,388,353 bytes |
Introducing the pretrained model reduced the compressed data by 4,966,036 bytes. The decoder ZIP, meanwhile, grew by 4,375,320 bytes, for a net reduction of 590,716 bytes. That is an improvement of about 0.56% over the first version's total size. The gain is large when looking at the compressed data alone, but shrinks once the model's storage cost is included.
The research team also compared configurations that store more pretrained weights. "L6," which uses a six-layer pretrained model, shrank the compressed data to 99.173 MB, about 1.79 MB smaller than C2. But L6's full set of decoding files reached 10.945 MB, making its total 110.118 MB, about 4.73 MB larger than C2. He chose C2 based on total size, including the decoding files, rather than on compressed data alone.
Comparisons were also run on enwik8, the first 100 million bytes of enwik9. In that test, C2's total size was 17.58 MB, while a configuration using online learning alone was smaller at 15.15 MB. On enwik9, with ten times as much data, C2 came out ahead. The fixed cost of storing a pretrained model is easier to recoup as the amount of data to be compressed grows. However, both datasets were extracted from the same Wikipedia dump, and the exact data size at which C2 becomes advantageous has not been identified.
Note that the MB figures in the research report are in decimal notation, and some are rounded.
About 7.5 days to decode: a wide gap between size reduction and speed
According to He, the first version of tufazip ranked second on the Large Text Compression Benchmark when it was submitted on September 7. Other methods subsequently set new records, and v2, which reduced total size further, stands at 13th in the overall ranking updated on November 19. Rankings are determined by total size including decoding files; compression and decompression times and memory use are not part of the evaluation. Tufa Labs says v2 achieved the smallest total size among the listed methods that use only a neural network to predict the next token.
Processing time, however, is extremely long. According to the README published by the developer on GitHub and the benchmark's per-version records, v2 took 631,548 seconds (about 7.3 days) to compress and 645,748 seconds (about 7.5 days) to decompress. The compression time does not include the time spent pretraining the initial model.
The benchmark's overall ranking table lists 647,595 seconds for compression and 630,136 seconds for decompression for v2, but these figures match the first version's numbers in the per-version records on the same page. Tom's Hardware's November 21 article also uses the figures from the README and the per-version records, 631,548 seconds and 645,748 seconds, for v2.
Running the published decoder requires a Linux x86-64 environment and exactly two NVIDIA B200 GPUs visible to the system. Besides the required runtime environment such as CUDA and Python, the GitHub README includes SHA-256 hashes of the compressed data and of the restored enwik9. This lets users who run the decoder verify that the restored data matches the original byte sequence exactly. Tufazip avoids distributing a large pretrained model, but because the decoder repeats the training process, it requires substantial computing resources and time.
The difference in speed becomes clearer when compared with another compression method. "DNA-4," which also targets enwik9, recorded a total size including decoding files of 105,554,360 bytes on the benchmark. Tufazip v2 is 166,007 bytes smaller, but DNA-4's decoding time was 9,517 seconds (about 2.64 hours). DNA-4 used an NVIDIA RTX 4090 and an Intel Xeon Platinum 8352Y; because the compression methods and execution environments differ, their processing performance cannot be compared directly. Even so, it shows that methods achieving nearly the same storage size can differ greatly in the time needed for decoding.
These results measure compression performance on a fixed Wikipedia text, and the total time and power consumption required for pretraining cannot be confirmed from the published materials. Open questions include whether the weights of the initial model can be made smaller while maintaining prediction performance, and whether the compression gain on different texts would exceed the added model size. Comparing processing time and computational cost with the input data and execution environment held constant would also allow use-case-specific judgments about whether to prioritize smaller storage or shorter decoding time.
