Photos of a lecture slide by Jie Tang, the Tsinghua University professor who is chief scientist at Z.ai (Zhipu), show the data behind the company’s GLM models coming from six sources, led by 1 trillion web pages and 500 million code repositories. The photos were posted on X by @bebuu_ai on 4 November, and the post had been viewed about 128,000 times by 5 November.
The slide is titled “GLM: six sources, six scaling problems”. It lists a web crawler running at about 10 billion pages a day. By our arithmetic, at that rate 1 trillion pages is about 100 days of crawling.
Business Pill · Data Quality
An evergreen Business Engineer explainer on data quality: why fewer, better training examples can beat more data. It teaches the general idea only and does not discuss Z.ai or GLM.
The key insight: A Jie Tang slide lists six data sources behind GLM, led by 1 trillion web pages and 500 million code repositories. The slide does not mention GLM-5.5; that link comes from the post that spread it.
The slide gives a scale and the key engineering for each source, as we read it from the photo:
Web: 1T pages. A crawler at ~10B pages a day, entropy-scored link discovery, an in-house parser and PB-scale template clustering.
Documents: 2B. GLM-OCR for scientific documents, with speculative decoding making it 4x faster.
Books: 50M. Quality selection, garbled-text detection and layout reordering.
Code: 500M repos. “Executable tasks with tests and issue descriptions”.
Video: 10B. “World knowledge and long sequences”.
Audio: 300M hours. Speech recognition plus iterative rewriting, into “trillions of tokens” of world knowledge.
A line under the table reads: “Each source is a unique scaling problem: crawler, filter, dedup, quality.” A second photographed slide is titled “Data” and subtitled “Scaling effective tokens”, with a scaling-law formula in which the data term is boxed.
What Was Posted on X
http://Z.ai chief scientist Jie Tang just revealed the training pipeline behind the upcoming GLM-5.5 (32B, 122B, and 256B) during a lecture at Tsinghua University today.
Instead of relying on public datasets, their pretraining pipeline pulls from six separate internal engines:
– Web: 1 trillion pages crawled at ~10B pages/day with PB-scale template clustering – Documents: 2 billion scientific PDFs parsed through GLM-OCR (4x faster with speculative decoding) – Books: 50 million books with layout reordering and garbled-text filtering – Code: 500 million repositories turned into executable tasks with real tests and issue descriptions – Video: 10 billion videos for long-sequence world knowledge – Audio: 300 million hours rewritten via ASR into trillions of tokens
The post that spread the slide, on 4 November. It says the slide shows the pipeline behind an upcoming GLM-5.5 in three sizes; the slide photo itself does not mention GLM-5.5 or model sizes, which appear on a separate graphic in the post.
An earlier Chinese-language recap of the same lecture with the same six figures. It describes them as the data recipe of past GLM models and says the latest mix is not public.
The English post that spread the slide says it shows “the training pipeline behind the upcoming GLM-5.5 (32B, 122B, and 256B)”. The slide itself does not mention GLM-5.5 or any model size; the sizes appear on a separate graphic in the post, whose origin the post does not state. An earlier Chinese-language recap, posted by @NFT_Chen and viewed about 63,000 times, describes the lecture as revealing the data recipe of past GLM models (“ไปฅๅพGLMๆจกๅ็ๆฐๆฎ้ ๆน”) and says the latest mix has not been made public.
Z.ai has not announced a model called GLM-5.5 in any source we read, and we found no Z.ai statement confirming the slide’s figures. The small footer text on the slide is blurry; as we read it, it says “Tsinghua, 2024 Intern Talk”, so the slide may be older material reused in the lecture.
Why the Data Side Matters
Z.ai is one of China’s listed model makers. Our 3 November piece found Anthropic made 63 times the revenue of Zhipu and MiniMax, and another tracked the two shedding $150 billion from their peaks ahead of a January share unlock. Open-weight rivals are shipping fast: Moonshot’s Kimi K3.1 Pro weights came out this week, and DeepSeek V5 is said to be moving its pretraining to Huawei Ascend chips.
The Structural Read
As we read the slide, code is treated differently from the other sources: 500 million repositories become “executable tasks with tests and issue descriptions”, which is training material that can be checked, not only text to read.
As we read the two accounts, the English post framed the figures as GLM-5.5’s pipeline, while the earlier Chinese recap called them past GLM models’ recipe; the slide on its own supports neither model name.
Jie Tang’s slide, as photographed
“Each source is a unique scaling problem: crawler, filter, dedup, quality.”
Three Implications
FOR Z.AI The figures come from a lecture slide, not a Z.ai release; Z.ai has not confirmed them in any source we read.
FOR READERS OF THE VIRAL POST The GLM-5.5 name and the 32B, 122B and 256B sizes are on a separate graphic in the post, not on the slide.
FOR OPEN-WEIGHT RIVALS Moonshot released Kimi K3.1 Pro weights this week, per our earlier coverage.
The framework says: “Each moat is built by a different data pipeline.” It names a Corpus Moat built by pretraining and a Verifier Moat built on domain-specific reward signals.
As we read it, the slide has both kinds of pipeline: web pages, books and documents on the corpus side, and code repositories turned into tasks with tests. The framework does not discuss Z.ai.
What Is Not Established
Which GLM model the slide describes, when it was made, and whether the figures are totals collected, used in training, or targets. The photos come from a third party, and the model sizes in the viral post are not on the slide.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
A Jie Tang slide shows GLM’s data coming from six sources, including 1 trillion web pages crawled at about 10 billion a day and 500 million code repositories turned into tested tasks. The slide does not tie these figures to GLM-5.5; that link comes from the post that spread it.
A note on sourcing. Slide figures are read from photos posted on X by @bebuu_ai on 4 November 2026; Z.ai has not confirmed them. The days-of-crawling figure is our arithmetic from the slide.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
