GLM-5.5 Pipeline Claim: 1 Trillion Web Pages, Six Data Engines, 32B/122B/256B: Claimed vs Verified

The short answer: nobody but one X account has said these numbers out loud, and Z.ai has not confirmed them. On November 4, 2026, the account @bebuu_ai posted that Z.ai chief scientist Jie Tang "just revealed" the training pipeline for the upcoming GLM-5.5 during a lecture at Tsinghua University, with model sizes of 32B, 122B and 256B. The post drew more than 230,000 views, thousands of likes and a wave of replies. It names no video, slides or transcript.

We checked the post, a follow-up article that repeats it, and Z.ai's published GLM-5 paper. This is the claimed-versus-verified breakdown.

Magnifier over a claim bubble marked with a green check, showing how to verify the GLM-5.5 data pipeline claim

What the post claims

According to the X post by @bebuu_ai, published at 6:48 PM on November 4, 2026, the GLM-5.5 pretraining pipeline "pulls from six separate internal engines" instead of public datasets:

The post also gives the upcoming sizes as 32B, 122B and 256B.

What is verified

A site called RuntimeWire reported the claim the next day. It says the X post is its only source, links no recording, and that there is "no official Z.ai announcement confirming the figures." It also notes that the shipped flagship is still GLM-5.3. That article is a relay, not a second source, so it does not raise the confidence level.

Do the numbers hang together?

We cannot confirm the figures, but we can test whether they are consistent with each other. This is our own arithmetic, not a Z.ai statement.

• Web: 10 billion pages a day for 100 days gives 1 trillion pages. The claim is self-consistent. But pages crawled are not tokens trained on. After deduplication and quality filters, usable text is a small share of a raw crawl. Compare Marin's open build log for its 535B model, where a raw pool of 25.6 trillion tokens shrank to 23.1 trillion after deduplication and benchmark decontamination; see our Marin Mochi 535B report.
• Audio: 300 million hours of speech at roughly 9,000 spoken words an hour is about 2.7 trillion words, or a few trillion tokens. "Trillions of tokens" fits that estimate.
• Documents and books: 2 billion PDFs and 50 million books are large but plausible for a lab with a dedicated OCR model. A book holds about 100,000 words, so 50 million books would be about 5 trillion words before filtering.
• Code: 500 million repositories is a very large number. The post does not say whether it counts forks, mirrors or empty repositories, so the figure is hard to interpret.
• Video: 10 billion videos is the hardest to judge. No unit of length or token count is given.

None of this proves the numbers. It only shows they are not obviously impossible, which is a low bar.

Why the model sizes are the weak point

The post says GLM-5.5 comes in 32B, 122B and 256B. Three things make us hold that loosely.

1. Possible expert sizes. One reply in the thread read the numbers as likely expert sizes, not model sizes. The post does not say which. The current flagship GLM-5.3 is reported by third parties at about 744B parameters, so a 256B top model would be a big shrink, while 256B active would be a big growth.
2. Earlier circulation. RuntimeWire says the sizes "previously circulated as internal checkpoints." If so, the post is not new information about sizes.
3. No model card. Without a model card, parameter counts are only a rumor.

For how Z.ai has handled model naming and release timing before, see our timeline of the GLM-5.3 open-weights delay, where a Hugging Face placeholder listed a date that passed without a release.

What we do know about Z.ai and GLM

Some context is firmly established from earlier coverage and primary documents:

• Jie Tang has run public requests for feedback before. In June 2026 he asked the community what GLM-5.3 must include and got thousands of replies, mostly asking for vision.
• GLM-5.2 shipped under an MIT license, as covered in our July 2026 adoption report.
• GLM-5.3 scored 60 on the Artificial Analysis Intelligence Index, tying Kimi K3 for the top open-weights spot at the time, on the same base as GLM-5.2.
• The GLM-5 technical report credits a team of more than 180 authors, lists Jie Tang among them, and describes a sparse attention method, asynchronous reinforcement learning infrastructure, and strong results on software engineering tasks.

Z.ai has a record of describing its data work in public talks, which is why the claim is believable. Believable is not the same as confirmed.

A sea of small books flowing into a single green vessel, illustrating how six data engines feed one pretraining run

What it would mean if true

If the lecture did say this, three things follow.

Data is the moat. A lab that crawls its own web, parses its own PDFs with its own OCR model, and builds its own code tasks does not depend on public datasets such as Common Crawl. That fits the reply that said GLM models are strong for their size because of the dataset, though that reply is opinion.

Code tasks with tests. Turning repositories into "executable tasks with real tests and issue descriptions" is the kind of data used for agentic coding training. It matches the agentic focus of the GLM-5 report.

Openness. Few labs describe their pipelines at all. Marin shows what a full disclosure looks like: sources, filters and dedup counts published ahead of release. A claim at this level of detail from a commercial lab would be notable, which is why a document should back it.

What the claim does not tell you

• No benchmark. The post says nothing about how GLM-5.5 performs.
• No license. We do not know if GLM-5.5 will be released with open weights; GLM-5.3's open-weights status has its own delayed history.
• No price or API plan.
• No training token count, which is the number that would let you compare it with other models.

How to check this yourself

1. Look for primary evidence. A video of the lecture, slides, a Tsinghua event page, or an official Z.ai post. None was visible when we checked.
2. Check Z.ai's own channels. The Z.ai website, its Hugging Face organization and its official X account are where a real GLM-5.5 announcement would appear first.
3. Read the model card when it exists. The data section should state token counts and sources. Compare each number with the post.
4. Watch for corrections. If the original poster edits or deletes the thread, or if Z.ai replies, that settles it faster than any analysis.

We will update this post if Z.ai or Jie Tang confirms or denies any figure.

Bottom line

A popular X post gives specific, internally consistent numbers for the GLM-5.5 data pipeline. It is a single, second-hand report of a lecture no one has shown evidence of, and Z.ai has said nothing. The honest status is "reported, unconfirmed." Do not quote the 1 trillion pages, 32B/122B/256B or per-engine figures as facts, and do not plan a deployment around the sizes.

Related reading

• Z.ai GLM-5.3 open-weights delay: the verified timeline
• Jie Tang asks what GLM-5.3 must include
• GLM-5.3 ties Kimi K3 on the Intelligence Index
• GLM-5.2: MIT open weights and Code Arena adoption
• Marin Mochi 535B: an open data pipeline, in full view
• Z.ai GLM legal model: 19 percent gated access, claimed vs verified