Tufa Labs open-sources tufazip, pure neural lossless compression takes second on long-running benchmark
Tufa Labs' Tommy He released and open-sourced tufazip, which turns a Transformer's next-character prediction into lossless compression: the model predicts a character probability distribution, which is then arithmetic-coded into a more compact representation that can be decoded back to the exact original bytes. On enwik9 (1GB of Wikipedia text), tufazip ranks second on the 20-plus-year-old Large Text Compression Benchmark—the first entry where a neural network is the sole predictor among the best results, as all higher-ranked entries are not pure neural approaches.
Confirmed
- Compresses enwik9 (1GB Wikipedia text) to 105.39MB, nearly 90% smaller, strictly lossless
- The model is a 163M-parameter Transformer and the only benchmark entry using a neural network as its sole predictor
- Compression took about 7 days on two B200 GPUs
- tufazip is open-source, released by Tommy He of Tufa Labs
- In principle, the better a language model predicts the next character, the higher the compression ratio; tufazip is an engineering realization of this idea
Why it matters
- It validates the view that neural networks are the best compressors: for the first time a pure neural method ranks among the best on a long-standing public benchmark, providing empirical support for the equivalence of LLMs and information theory
- Communities such as Reddit are already discussing local deployment: @Terminator857 noted discussions that neural compression could let a local Wikipedia copy fit in roughly 27GB, hinting at AI compression's storage potential
- However, the compute and time cost (7 days on dual B200s) shows the approach is still research-oriented, far from everyday practicality
2026-11-19 ~ 2026-11-20 · 5 related posts
Primary sources
- Compress Wikipedia to ~27GB locally: neural compression via tufazip — Terminator857 · 2026-11-19
- tufazip losslessly compresses 1GB of Wikipedia to 105.39MB, best neural-only entry on 20-year-old benchmark — philipvollet · 2026-11-19
- tufazip: Transformer-Based Lossless Compression Squeezes 1GB Wikipedia to 105MB — aigclink · 2026-11-20
2 near-duplicate retellings: aigclink · adnan_hashmi