GLM-5.5's six data engines rest on one X post and a lecture nobody has seen

On 4 November an account called @bebuu_ai posted that Z.ai chief scientist Jie Tang "just revealed the training pipeline behind the upcoming GLM-5.5 (32B, 122B, and 256B)" in a lecture at Tsinghua University. The post drew more than 230,000 views, thousands of likes and a large Reddit thread. No recording, event page or statement from Z.ai has turned up. The honest status is reported, unconfirmed.

What the post claims

The thread describes six internal data engines in place of public datasets:

• Web: 1 trillion pages crawled, about 10 billion a day
• Documents and books: 2 billion scientific PDFs parsed with GLM-OCR, said to run 4x faster with speculative decoding, plus 50 million books
• Code, video, audio: 500 million repositories converted into executable tasks, 10 billion videos, 300 million hours of speech

Each figure is the post's, not Z.ai's, and the post cites no document behind them.

What could be checked

Two things survive contact. The post exists, is dated, and drew the engagement claimed for it. And Tang's role is real: Z.ai co-founder and chief scientist, and an author on the GLM-5 technical report, whose abstract describes sparse attention, asynchronous RL and benchmark results rather than a data pipeline.

The rest failed. The fact-check published on 6 November found no event page, no recording, no attendee account and no Z.ai confirmation or denial of any number. RuntimeWire, which relayed it the next day, names the post as its only source and notes there is "no official Z.ai announcement confirming the figures."

Why the sizes are the weak point

32B, 122B and 256B are the quotable part and the least supported. They were already in circulation: a 1 November write-up on kie.ai, relaying the Lumina account, described them as internal checkpoints trained to merge into a base model, not products. The thread split on exactly that, some asking why a 32B was not shipping separately, others answering that the numbers are internal, citing a follow-up post. Scale is why it matters: GLM-5.3, the model Z.ai ships, is a 744B mixture-of-experts with 40B active, as we covered in August. Read as a flagship, 256B would be a heavy shrink.

What would settle it

Less than a vendor confirmation. A recording or slide set from the talk, a dated event page, or one attendee describing the session would move it from relayed to reported. Failing that, a Z.ai post, model card or paper with a data section and token counts would do it; a correction or deletion of the thread would end it. Internal consistency proves nothing: 10 billion pages a day reaches 1 trillion in a hundred days, which validates the arithmetic, not the crawl.