Z.ai

Zhipu / Z.ai's flagship coding-first agentic model and successor to GLM-5.1 — a ~753B MoE with DeepSeek Sparse Attention (DSA) and the highest-scoring…

754B parameter MoE built for long-horizon agentic engineering. Designed to sustain optimization over hundreds of rounds and thousands of tool calls wi…

Z.ai's first native multimodal agent foundation model, built for vision-based coding and agentic workflows. Natively processes images, videos, design…

The sparse-attention efficiency technique behind GLM-5.2 (productized there as IndexShare; the paper calls it IndexCache). DeepSeek Sparse Attention (…

Most powerful Z.ai model family. GLM-5 is 744B total params (44B active) via MoE with 256 experts. Hybrid Attention and Multi-Token Prediction. First…

Introduced the training method that reduced GLM-5's hallucination rate by 35%.

100,000 complex multi-step reasoning trajectories used to train the GLM-5 "Thinking" mode.

Optimized for long-horizon planning and reducing hallucination.

Released with MIT License. 355B MoE model, premier open-weight choice for enterprise self-hosting.

Open-source vision-language model (106B total, 12B active via MoE). Supports 64K tokens for multi-image and video inputs. SOTA on 42 public VL benchma…

First "Z.ai" branded release. 355B Mixture-of-Experts model designed for Agentic Engineering. Claimed to be cheaper to use than DeepSeek.

Benchmark for evaluating an agent's ability to browse the web and synthesize diverse sources.

Collaborative library optimized for running massive MoE models (like GLM-4.6/5) on consumer GPUs.

Latest text-to-image generator, capable of cinematic detail and accurate text rendering within images. Utilizes a Diffusion Transformer (DiT) architec…

Benchmark for evaluating multi-step and constrained function calling under long-context scenarios. Tests LLM tool-use capabilities with complex, real-…

Intelligent and human-like end-to-end spoken chatbot.

"Computer-Use" agent capable of navigating smartphone apps and web browsers via natural language commands.

Next-generation vision-language model. CogVLM2-Video introduced multi-frame time localization for precise video understanding.

Text-to-video diffusion model with expert transformer architecture.

Native multimodal version of the GLM-4 architecture.

Theoretical analysis of emergent abilities in language models through the lens of pre-training loss, providing a unified framework for understanding w…

Introduces "Relay Diffusion," a strategy that breaks image generation into a low-resolution stage followed by a super-resolution process. CogView3 out…

Major flagship upgrade featuring "All Tools" integration (web search, code interpreter, and vision).

Self-reflective instruction annotated dataset for training scientific language models. Introduces SciInstruct for math, physics, and chemistry reasoni…

Powerful vision-language model using visual expert modules within pretrained language models.

Comprehensive benchmark for long-context understanding in LLMs.

Iterative improvements over ChatGLM-6B in context length, tool use, and reasoning. ChatGLM2 released June 2023, ChatGLM3 released Oct 2023.

Web-enhanced question answering system with human preferences. Combines web search, LLM-based filtering, and answer generation with human preference a…

High-efficiency open-source conversational model that sparked a wave of local LLM deployment in China.

Dedicated series for software engineering (versions 1-4), now fully integrated into AutoGLM for autonomous repository management.

Landmark bilingual (English/Chinese) model among the first in China to rival GPT-3.

A hierarchical transformer-based text-to-image model (6B parameters) that introduced the Cross-modal General Language Model (CogLM) and Local Parallel…

The first large-scale generative model for text-to-image synthesis in the general domain for Chinese. CogView is a 4-billion-parameter Transformer usi…

The GLM paper (ACL 2022) is a seminal work that unified the three major pretraining paradigms — autoencoding (BERT), autoregressive (GPT), and e…

Prompt tuning methods for pre-trained language models. P-Tuning v1 ("GPT Understands, Too") introduced trainable continuous prompts. P-Tuning v2 showe…