According to an X post published on November 4, Z.ai chief scientist Jie Tang reportedly outlined the training pipeline for the upcoming GLM-5.5 during a lecture at Tsinghua University. The company has neither confirmed nor denied the details, and the model has not been officially announced.

Psst!Add AlexTech.ai asPreferred Source on Google

The post says Z.ai would not rely on public datasets for pretraining, but on six internal engines, each dedicated to a different type of content. The first covers the web: a corpus of one trillion pages, crawled at roughly 10 billion pages per day and organised through template clustering at petabyte scale. The second covers documents: 2 billion scientific PDFs parsed with GLM-OCR, the company's optical recognition system, which the post says runs four times faster thanks to speculative decoding.

official image of GLM-5.5

Image: bebuu_ai — https://x.com/bebuu_ai/status/2118052769856888892

The other four streams complete the picture. Books would total 50 million, processed for layout reordering and filtering of garbled text. Code would include 500 million repositories, turned into executable tasks with real tests and issue descriptions. Video would reach 10 billion clips, used to build world knowledge over long sequences. Audio would amount to 300 million hours, rewritten through speech recognition into trillions of tokens.

The figures come from a single source, the post attributing them to Tang. The material circulating online includes no video, slides or transcript of the lecture, so the numbers remain unverified. RuntimeWire notes that no official Z.ai announcement confirms them.

Context helps gauge the claim. On GitHub, Z.ai describes GLM-OCR as a multimodal OCR model for complex document understanding, built on the GLM-V encoder-decoder architecture and using a Multi-Token Prediction loss. The company launched GLM-5 on February 12, 2026, targeting complex systems engineering and long-horizon agentic tasks. According to Tang's X profile, that generation scaled from 355 billion parameters with 32 billion active to 744 billion with 40 billion active, and pretraining grew from 23 to 28.5 trillion tokens. For GLM-5.3, covered earlier on AlexTech, the approach had been to scale post-training alone, without changing the architecture.

The sizes attributed to GLM-5.5, 32B, 122B and 256B, have circulated online for several days. According to the posts gathered by the sources, these would be internal models to be merged into the next base model, not three separate releases. Timing remains unclear.

For observers, the key point is that competition over training data is becoming a distinct battleground from architecture. If Z.ai confirmed even part of the scale described, proprietary data would become a decisive factor for its next flagship model. For now, the material remains to be verified.