IndexMapAnti-DoomscrollAbout
Model FamilyLLMCross-Modal

GLM (ChatGLM & GLM-4 / GLM-5)

Also known as: GLM · ChatGLM · GLM-4 · GLM-4.5 · GLM-5 · GLM-130B

Plain English

Zhipu AI and Tsinghua KEG's flagship family of open-weight and API language models—evolving from the early ChatGLM-6B local model into the large-scale GLM-4.5 and GLM-5 Mixture-of-Experts coding and reasoning models.

Technical Overview

Originally formulated around autoregressive blank infilling (GLM-130B, ChatGLM), the lineage transitioned to decoder-only Transformers with 2D/long-context RoPE in GLM-4, and to high-sparsity Mixture-of-Experts architectures with hybrid thinking/non-thinking modes and sparse attention in GLM-4.5 (355B/A32B) and GLM-5 (744B/A40B).

Timeline (5)

RSS
  1. 2022

    Paper

    Aohan Zeng et al. at Tsinghua KEG and Zhipu AI published GLM-130B: An Open Bilingual Pre-trained Model, showing that a bidirectional autoregressive blank-infilling objective and DeepNorm stabilization scaled cleanly to 130B parameters with lossless INT4 weight quantization.

  2. 2023

    Release

    Tsinghua and Zhipu AI released ChatGLM-6B, an open-weight 6.2B bilingual dialogue model runnable in 6 GB of VRAM via INT4 quantization that catalyzed early Chinese local LLM deployment and LoRA fine-tuning.

  3. 2024

  4. 2025

  5. 2026

    Release

    Z.ai released GLM-5 under the MIT license, scaling the family to a 744B-parameter (40B active) MoE architecture with DeepSeek Sparse Attention (DSA) and asynchronous reinforcement learning (slime) across 28.5T pretraining tokens.

See also (15)

Open full graph
10 more connections