GLM (General Language Model) is the model family of Zhipu AI, distinctive for having entered open frontier publishing before the ChatGPT era and for maintaining continuous open-weight releases through the agentic-model race of 2025-2026.1

Family tree

Generations

GLM-130B (2022) used an autoregressive blank-infilling objective distinct from both BERT-style masking and pure next-token prediction, and its documented open release was nearly unique for its scale at the time. ChatGLM-6B (2023) became a fixture of consumer-hardware deployment in China. The GLM-4 generation converged on mainstream decoder-only design; GLM-4.5 (2025) adopted mixture-of-experts scaling and agentic post-training with verifiable-reward RL, positioning the line against DeepSeek, Qwen, and Kimi in the open-weight coding race, with results on MMLU-class and agentic suites reported by the developer. The GLM-5 generation (2026) is covered under the epistemic rules for recent releases.12

Licensing and adjacent lines

Recent GLM flagships ship under permissive licenses through Hugging Face, and the family travels with siblings: the CodeGeeX code models, the CogView and CogVideoX generative-media lines, and agent-focused variants. The lab's own family retrospective traces the arc from blank-infilling pretraining to the all-tools assistant design, an unusually candid published account of a frontier line's architectural pivots.3 Within China's open-weight race the GLM line is the longest-running continuous lineage, predating both the DeepSeek and Kimi families by over a year.3

See also

References


  1. Zeng, A., et al., "GLM-130B: An Open Bilingual Pre-trained Model," arXiv:2210.02414, October 2022. ↩↩

  2. Zhipu AI / Z.ai, GLM-4.x release materials and technical reports, 2024-2026. ↩

  3. GLM Team, "ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools," arXiv:2406.12793, June 2024. ↩↩