Tsinghua KEG and Zhipu AI released GLM-130B, an open-weight bilingual English-Chinese 130B language model capable of INT4 weight Quantization without perplexity degradation on a single 4x RTX 3090 or 8x V100 node.
Zhipu AI (Z.ai)
Also known as: Zhipu · Z.ai · THUDM
Plain English
One of China's leading frontier AI labs—spun out of Tsinghua University's KEG lab and creator of the open-weight GLM language models, CogVideoX video generators, CogView image models, and AutoGLM GUI agents.
Technical Overview
Operating internationally as Z.ai and publishing on GitHub/Hugging Face under THUDM and zai-org, Zhipu AI builds the GLM language and agentic MoE lineage (ChatGLM, GLM-4, GLM-4.5, GLM-5), full-3D-attention CogVideoX video DiTs, and phone/browser GUI agent harnesses (AutoGLM).
Timeline (4)
2022
Release2024
ReleaseZhipu AI open-sourced CogVideoX (
2Band5B, followed byCogVideoX1.5-5Bin November 2024), establishing full 3D spatiotemporal attention over a 3D causal VAE as a core open-weight video generation architecture.2025
ReleaseRebranded internationally as Z.ai, Zhipu released GLM-4.5 (355B total / 32B active MoE) and
GLM-4.5-Air(106B total / 12B active) as open-weight models unifying hybrid thinking reasoning, coding, and agentic tool use.2026
ReleaseFollowing its January 2026 Hong Kong Stock Exchange IPO, Z.ai released GLM-5 under the MIT license—scaling to a 744B-parameter (40B active) sparse-attention MoE trained on 28.5T tokens for long-horizon software engineering and agent workflows.
See also (14)
Open full graphZhipu AI (Z.ai)
9 more connections
- StepFun (Step-Video & Step-Audio)Incoming · Alternative to
- Tencent HunyuanIncoming · Alternative to
- Agent HarnessOutgoing · Mentions
- Diffusion Transformer (DiT)Outgoing · Mentions
- INT4Outgoing · Mentions
- Latent Space & VAEOutgoing · Mentions
- Mixture of Experts (MoE)Outgoing · Mentions
- QuantizationOutgoing · Mentions
- Video Generation Models (Sora, Veo, Kling, Wan)Outgoing · MentionsIncoming · Mentions