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

Why it matters

2026-11-19 ~ 2026-11-20 · 5 related posts

Primary sources

2 near-duplicate retellings: aigclink · adnan_hashmi