# Z.ai's Jie Tang outlines training data pipeline for upcoming GLM-5.5 models

> Jie Tang described the pretraining pipeline for GLM-5.5 variants labeled 32B, 122B and 256B, according to a post relaying a Tsinghua lecture.

- Status: reported (One outlet, not independently verified yet.)
- Topic: Models
- Published: 2026-11-06T10:36:15+00:00
- Updated: 2026-11-06T10:36:15+00:00
- Web page: https://ai-news.ikangai.com/story/z-ai-s-jie-tang-outlines-training-data-pipeline-for-upcoming-glm-5-5-models/
- Seen in: Latent Space
- Change: other (Other)
- Applies to: GLM-5.5

Z.ai chief scientist Jie Tang has outlined the training data pipeline for upcoming GLM-5.5 models in three sizes: 32B, 122B and 256B. The outline comes from a post on X by bebuu_ai that relays a Tsinghua lecture. The pipeline is described as drawing primarily on internal data, with six sources: web crawling of about 1T pages, which the post pairs with a rate of about 10B pages per day; scientific PDF extraction using GLM-OCR; books; executable code tasks mined from code repositories; video-derived data for long-context and world-knowledge training; and ASR-transcribed audio, meaning speech converted to text by automatic speech recognition, totaling hundreds of millions of hours.

Commenters in a discussion thread about the post were positive about the parameter sizes, especially the smaller 32B model. One technical reading raised in that thread is that the 32B, 122B and 256B models may not be standalone releases. They could instead be component checkpoints meant to be merged into a larger system of roughly 1T parameters. The thread presented this as speculation, and nothing in the relayed account says Z.ai has stated it.

A second open question concerns architecture. Commenters asked whether the 32B variant will use a mixture-of-experts design, in which each token activates only a subset of the model's parameters. If it does, the smaller model would be more compute-efficient at inference than a dense 32B model, which uses all of its parameters for every token. The post does not say whether any of the GLM-5.5 variants use this design.

## Why it matters

The description points to heavy reliance on proprietary data pipelines. That matters for teams weighing Chinese open models against alternatives, though the details come from a secondhand account.

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Z.ai's Jie Tang outlines training data pipeline for upcoming GLM-5.5 models", 2026-11-06, https://ai-news.ikangai.com/story/z-ai-s-jie-tang-outlines-training-data-pipeline-for-upcoming-glm-5-5-models/
