tufazip compresses 1B bytes of Wikipedia to 105MB, ranks #2 on 20-year benchmark
adnan_hashmi · x · 2026-11-20
Tommy He of Tufa Labs released tufazip, taking #2 on the 20-year-old Large Text Compression Benchmark — the best entry using a neural network as its sole prediction model. It compresses one billion bytes of Wikipedia into 105,388,353 bytes (105.39 MB), including offline-trained weights and the decoder.
Key details:
- The final system is a 12-layer, 163M-parameter Transformer initialized from a 34M-parameter offline model quantized to just 4.33 MB.
- Two offline-trained layers are copied into twelve independent layers that keep learning as the file is compressed; the decoder repeats the same learning process, so full model weights never need to be transferred.
- The writeup also connects compression to intelligence via Legg & Hutter's and Chollet's definitions, echoing the Hutter Prize tradition.
More from Infra
- Microsoft's USB4 RDMA lets mini PCs pool memory to run Qwen3-235B — JFPuget · 2026-11-22
- LTX 2.5 keeps crashing on AMD 9070 XT with driver timeouts while MiniMax H3 runs fine — ZestycloseReading504 · 2026-11-22
- FT: AI server suppliers raise prices up to 45% months after deals are signed — nordicinst · 2026-11-22
- China Installed 59% of New Industrial Robots in 2026 Report, ~354,000 Units — ingliguori · 2026-11-22
- MLX over Vulkan hits 124 tok/s on RTX 5060 Ti, beating llama.cpp CUDA by 25% — SumitGup · 2026-11-22
- Intruder swapped Prime Intellect's hosted DeepSeek for an abliterated copy for 53 hours — mawcopolow · 2026-11-22