GLM · Zhipu AI model

Also known as: ChatGLM · Zhipu GLM

On this page 7

GLM is the family of large language models developed by Zhipu AI, a Beijing company spun out of Tsinghua University research, and in the field of AI models & assistants the name refers to this Zhipu AI model line, its open-weight releases and the assistant products built on it. The acronym originally stood for General Language Model, the pretraining method described by the Tsinghua group in 2021.

The family includes the bilingual ChatGLM chat models, the GLM-4 generation, vision variants and the GLM-4.5 mixture-of-experts series aimed at reasoning, coding and agent tasks. Many versions are released with open weights under permissive licenses, which has made GLM one of the most widely deployed Chinese model families alongside Qwen, DeepSeek and Kimi. Outside China, the company presents its platform under the Z.ai brand.

Origins of the GLM pretraining method

The original GLM introduced autoregressive blank infilling as a pretraining objective. Spans of text are removed from a passage, and the model learns to regenerate each missing span token by token while attending bidirectionally to the remaining context. The goal was a single architecture that handles understanding and generation, tasks that were then split between encoder models such as BERT and decoder models such as GPT.

This objective shaped the first large release, GLM-130B in 2022, a bilingual English and Chinese model with 130 billion parameters that was notable for supporting INT4 quantization on a single server of consumer-class GPUs. Later generations moved toward the standard decoder-only design used across the industry, keeping the GLM name as a brand rather than a strict description of the training objective.

Main generations of the GLM family

The GLM line has evolved through several generations, each extending context length, multilingual quality and tool use. The table summarizes the principal milestones without listing every intermediate checkpoint.

ReleaseApproximate periodKey characteristics
GLM-130B2022Bilingual dense model, open research weights
ChatGLM-6BEarly 2023Small open chat model runnable on one consumer GPU
ChatGLM2 and ChatGLM32023Longer context, function calling, code interpreter support
GLM-4 and GLM-4-9B2024Stronger multilingual ability, 128K context, open 9B variant
GLM-4.5 and GLM-4.5-AirMid-2025Mixture-of-experts, hybrid reasoning, agent and coding focus
GLM-4.6 and laterLate 2025 onwardLonger context and refined coding and agent performance
Principal GLM releases and their focus

Release names, parameter counts and licenses change between versions, so practitioners should verify the exact model card before deployment. Older ChatGLM licenses carried usage restrictions, while the GLM-4.5 generation was published under the MIT license.

Architecture and capabilities of current GLM models

Recent flagship GLM models use a mixture-of-experts (MoE) transformer, in which a router activates only a subset of expert feed-forward networks for each token. GLM-4.5 is reported at about 355 billion total parameters with 32 billion active, and GLM-4.5-Air at about 106 billion total with 12 billion active, which lowers inference cost relative to a dense model of similar quality.

These models offer a hybrid reasoning mode: a thinking mode that produces an extended chain of reasoning for mathematics, coding and planning, and a direct mode for fast replies. They support native function calling, structured outputs and long contexts, which makes them suitable as the engine of coding assistants and autonomous agents.

VariantModalityPrimary purpose
GLM text modelsTextChat, reasoning, coding and agents
GLM-4V and GLM-4.5VImage and textVisual question answering and document reading
CogVLM and CogAgentImage and GUIVisual grounding and interface automation
CodeGeeXCodeCode completion in development tools
CogView and CogVideoXImage and video generationText-to-image and text-to-video
Main GLM variants by modality and purpose

How GLM compares with other model families

GLM competes most directly with other open-weight model families from Chinese and Western developers. Its distinguishing traits are strong Chinese-English bilingual performance, an early emphasis on agentic coding and a broad portfolio that spans text, vision and generation models.

FamilyDeveloperNotable emphasis
GLMZhipu AIBilingual chat, agents, coding, vision variants
QwenAlibabaVery wide range of sizes and modalities
DeepSeekDeepSeekEfficient MoE training and reasoning models
KimiMoonshot AILong context and agentic tool use
LlamaMetaWidely adopted open baseline outside China
GLM compared with related open model families

Benchmark rankings among these families shift with every release, and results depend heavily on the task, the prompt format and whether reasoning mode is enabled. Comparisons should therefore rest on evaluation with the intended workload rather than on leaderboard positions alone.

Deployment options and practical considerations

GLM models are available through three main routes: the hosted API and chat assistant operated by Zhipu AI, third-party inference providers, and self-hosting of the open weights with serving engines such as vLLM or SGLang. Self-hosting gives full control over data location, but the largest MoE versions require multiple high-memory GPUs even with quantization.

SituationRecommended routeReason
Prototype or low volumeHosted APINo infrastructure, pay per token
Strict data residencySelf-hosted open weightsData stays on controlled hardware
Limited GPU budgetGLM-4.5-Air or a smaller variantLower active parameters and memory
Edge or laptop useQuantized 9B-class modelRuns on a single consumer GPU
Choosing a GLM deployment route
  • Regulatory context matters: models trained and served in China follow local content rules, which can affect answers on sensitive topics.
  • Export-control and sanctions screening may apply to commercial relationships with the developer in some jurisdictions.
  • License terms differ by version, so each checkpoint must be reviewed before commercial use.

In everyday language

GLM is a set of AI models made by a Chinese company called Zhipu AI. You can think of it as a Chinese counterpart to the models behind well-known chat assistants. It can answer questions, write and fix code, read documents and, in some versions, understand pictures.

What makes GLM stand out is that many versions are free to download and run on your own computers. A company that does not want to send its data elsewhere can install GLM on its own servers. It also handles Chinese and English equally well, which is useful for anyone working across both languages.

Seen in the real world

A mid-sized software firm that serves clients in both mainland China and Europe needs a coding assistant that keeps source code inside its own data center. The engineering team evaluates several open-weight models on its internal repository and selects GLM-4.5-Air, served with vLLM on a node of four 80 GB GPUs using 8-bit weights.

Developers connect the model to their editor and to an agent tool that can run tests and read files. On a set of 200 internal bug-fix tasks, the team records the share solved without human edits and compares thinking mode with direct mode, finding that thinking mode solves noticeably more tasks at roughly double the response time.

The firm keeps thinking mode for complex refactoring and uses direct mode for quick completions. The outcome is a bilingual coding assistant with no code leaving the company, at a hardware cost the team judged acceptable for about 150 developers.