Physix Frontier badgePhysix Frontier
PublishChinese | EN

Physix Frontier · News Briefing Card (Hacker News · Nov 6, 2026)

Zhipu reveals GLM-5.5 training data pipeline with 1 trillion web pages

KEY FACTS

  • Zhipu chief scientist Tang Jie disclosed the GLM-5.5 training data pipeline during a lecture at Tsinghua University.
  • GLM-5.5 will ship in three parameter versions: 32B, 122B, and 256B.
  • The pretraining data comes from six in-house engines rather than public datasets.
  • The web data totals 1 trillion pages, crawled at roughly 10 billion pages per day.
  • The corpus also includes 2 billion papers, 50 million books, 500 million code repositories, 10 billion videos, and 300 million hours of audio.

KEY DATA

1 trillion pagesWeb data volume
~10 billion pages/dayCrawl speed
2 billionScientific PDFs
50 millionBooks

PHYSIX OBSERVATION

Competition among large models is shifting from parameter counts to data pipelines. By treating its six in-house engines as a moat, Zhipu is effectively declaring that public corpora are no longer sufficient. For users, the ceiling on model capability now depends more on the quality of data cleaning and structuring; but the compliance costs and copyright risks of in-house crawling and parsing will also become a barrier that peers cannot easily sidestep.

Source: Hacker News report