BPBench finds Chinese open-weight models dominate text compression scores

ctnzr · x · 2026-07-25

The thread discusses BPBench, a benchmark that treats a pretrained LLM as a compressor: the better the model understands language, the fewer bits it needs to encode unseen text.

It measures models on fresh human writing published after their training cutoff, using:

The author says that among open-weight models, the best compressors are mostly Chinese models, with NVIDIA’s Nemotron as the notable Western outlier in the frontier cluster. The post also notes that commercial APIs make direct compression measurement hard because they block logprobs, and sampling-based workarounds are too noisy to recover the tails of the distribution. This also makes KL/log-prob distillation against closed APIs difficult.

Original post →

More from Research

Research channel →