The GLM family

GLM: Z.ai's open model family, explained

Name: GLM model family comparison
Creator: Ox Alpha

GLM (General Language Model) is the line of AI models built by Z.ai, formerly Zhipu AI. It runs from the small open ChatGLM models to the frontier-scale GLM-5 generation: GLM-5, GLM-5.1, GLM-5.2 and the current GLM-5.3. Here is the whole family, honestly, and how to run each one.

Open weight, MIT licensed at frontier scale

GLM stands for General Language Model, an architecture first published out of Tsinghua University in 2021. The company behind it, Z.ai (formerly Zhipu AI), has made GLM one of the most openly available frontier model families: the recent releases are published under a plain MIT license, so you can download, self-host, fine-tune, and commercially deploy them.

The family is positioned around agentic, reasoning, and coding work. Below is the modern lineup compared side by side, the full lineage from 2021 to today, and every way to run a GLM model. This is an independent reference, not affiliated with Z.ai. Last verified October 4, 2026.

2021

First GLM paper

MIT

License on recent models

Context (GLM-5.2, GLM-5.3)

744B

GLM-5 generation parameters

The lineup

The modern GLM models, compared

The recent, fully-specced members of the family, newest first. Every row is provider-published and every model has open weights; all but GLM-5.3 are MIT. The older models are in the timeline below.

Model	Released	Parameters	Context	Modalities	License
GLM-5.3 Flash Cheap, fast, 1M-context multimodal (the model behind Ox Alpha) source	Aug 2026	320B total / 18B active	1M	Text + image + video	MIT
GLM-5.3 Z.ai's current flagship: the strongest GLM for coding and agents, 1M context source	Aug 2026	753B total (GLM-5.2 base)	1M	Text	GLM-5.3 License
GLM-5.2 The newest GLM flagship under plain MIT: 1M context, long-horizon coding source	Jun 2026	744B / 40B active	1M	Text	MIT
GLM-5.1 GLM-5 retrained for long-running agentic engineering, 200K context source	Apr 2026	744B / 40B active	200K	Text	MIT
GLM-5 The 744B model that opened the GLM-5 generation, 200K context source	Feb 2026	744B / 40B active	200K	Text	MIT
GLM-4.7-Flash A 30B-A3B model for one machine, free on Z.ai's API source	Jan 2026	30B / ~3B active	200K	Text	MIT
GLM-4.7 The last GLM-4.5-architecture flagship: a 355B coding model, 200K context source	Dec 2025	355B / 32B active	200K	Text	MIT
GLM-4.6V Vision model with native multimodal function calling source	Dec 2025	106B total / 12B active	128K	Text + vision (images, video, documents)	MIT
GLM-4.6 Real-world coding and long-context agents (the prior open flagship) source	Sep 2025	~355B total / 32B active	200K	Text	MIT
GLM-4.5V Multimodal document, screen, and video understanding source	Aug 2025	106B total / 12B active	Not disclosed	Text + vision (images, video, PDFs, GUI)	MIT
GLM-4.5 Open-weight agentic, reasoning, and coding flagship source	Jul 2025	355B total / 32B active	128K	Text	MIT
GLM-4.5-Air Compact, cheaper-to-run sibling of GLM-4.5 source	Jul 2025	106B total / 12B active	128K	Text	MIT

Do not confuse GLM-5.3 Flash with the larger GLM-5.3 (a separate, more capable model released under a different, non-MIT license). Flash is the smaller, cheaper, MIT-licensed model that ran as Ox Alpha.

GLM-5.3 GLM-5.2 GLM-5.1 GLM-5 GLM-4.7-Flash GLM-4.7 GLM-5.3 Flash

What each GLM model costs Compare GLM to GPT, Claude, and Gemini

The lineage

From ChatGLM to GLM-5.3

How the family grew from a 2021 research architecture into an open, frontier-scale model line. Each step links to its source.

March 18, 2021
The GLM architecture is introduced in the paper 'GLM: General Language Model Pretraining with Autoregressive Blank Infilling', out of Tsinghua University's KEG lab. GLM stands for General Language Model.
arxiv.org
August 2022
GLM-130B, a 130B open bilingual (Chinese and English) base model, is released and later presented at ICLR 2023. One of the first openly available models at the 100B scale.
github.com
March 2023
ChatGLM-6B, a small open bilingual chat model, is released and becomes a breakout in the open-source community. ChatGLM2-6B and ChatGLM3-6B follow later in 2023.
github.com
June 18, 2024
The GLM-4 generation arrives: a proprietary API flagship (GLM-4, later GLM-4-Plus) plus the open GLM-4-9B (with an 8K base and a 1M-context variant) and the vision model GLM-4V-9B.
huggingface.co
July 2025
GLM-4.5 and GLM-4.5-Air ship as open-weight Mixture-of-Experts models under the MIT license (355B/32B and 106B/12B), branded as 'Agentic, Reasoning, and Coding' foundation models. Independent coverage rates GLM-4.5 among the strongest open-weight models at launch.
huggingface.co
August 2025
GLM-4.5V, a multimodal vision-language model built on GLM-4.5-Air, is released under MIT, handling images, video, PDFs, and on-screen content.
huggingface.co
September 30, 2025
GLM-4.6 lands under MIT with a 200K context window (up from 128K), stronger real-world coding and agentic tool use, and roughly 15% fewer tokens per task than GLM-4.5.
huggingface.co
December 2025
GLM-4.6V (106B) adds native multimodal function calling to the vision line, and GLM-4.7 (December 22, MIT) brings large agentic-coding gains over GLM-4.6 plus Preserved and Turn-level Thinking. Its 30B sibling GLM-4.7-Flash follows on January 19, 2026.
huggingface.co
February 11, 2026
GLM-5 doubles the family's scale to 744B parameters (40B active), trained on 28.5T tokens with DeepSeek Sparse Attention, under MIT. Z.ai had previewed it anonymously on OpenRouter as 'Pony Alpha'.
huggingface.co
March to April 2026
GLM-5 Turbo, a faster API-only model for agent workflows, then GLM-5V-Turbo, Z.ai's first native multimodal agent model. In April, GLM-5.1 retrains GLM-5 for long-horizon agentic engineering and tops Z.ai's SWE-Bench Pro comparison.
huggingface.co
June 16, 2026
GLM-5.2 extends the context window to a full 1M tokens and lifts coding sharply (Terminal Bench 2.1 rises from 63.5 to 81.0), still under MIT.
huggingface.co
August 14, 2026
GLM-5.3, the GLM-5.2 base with much more post-training, becomes Z.ai's flagship. Its open weights follow on August 28 under a new GLM-5.3 License.
huggingface.co
August 2026
GLM-5.3 Flash (320B/18B, 1M context, MIT) debuts anonymously as the stealth 'Ox Alpha' model, then is revealed as Z.ai's and open-weighted, the model this site is built around.
openrouter.ai
September 2026
Z.ai adds faster-serving variants: GLM-5.3 FlashX (up to 200 tokens per second) and GLM-5.3 Prime (1.5 to 2 times GLM-5.3's output throughput).
openrouter.ai

Access

Every way to run a GLM model

Use Z.ai's own API or its GLM Coding Plan, run it through a third-party host, or self-host the open weights. The recent models are MIT, so nothing locks you in.

GLM API pricing, every model and host Drop-in API settings for GLM-5.3 Flash The GLM Coding Plan, explained About Z.ai, the maker

Z.ai's own API: Call GLM models directly from Z.ai's OpenAI-compatible API at api.z.ai (base https://api.z.ai/api/paas/v4/).source Confirmed
OpenRouter and other hosts: GLM models are served by many third-party hosts. GLM-5.3 Flash is z-ai/glm-5.3-flash on OpenRouter, where Z.ai itself charges its list price of $0.15 / $0.50 per 1M input / output tokens and some hosts charge less. The GLM API pricing page compares every model and host.source Confirmed
Self-host the weights: Most GLM models publish their weights on Hugging Face under zai-org: GLM-4.5 through GLM-4.7, GLM-5, GLM-5.1, GLM-5.2 and GLM-5.3 Flash under MIT, and GLM-5.3 under its own GLM-5.3 License. Free to download, self-host, fine-tune, and use commercially.source Confirmed
GLM Coding Plan: Z.ai's monthly subscription for using GLM-5.3 and GLM-5.3-Flash inside coding agents (Claude Code, Cline, OpenCode, and more) instead of paying per token. Three tiers (Lite, Pro, Max), each a credit allowance with a rolling 5-hour limit and a weekly limit; requests for GLM-5.2 and GLM-5.1 are routed to GLM-5.3. Current prices at z.ai/subscribe.source Confirmed

FAQ

Straight answers about GLM

What is GLM?

GLM (General Language Model) is the family of AI models built by Z.ai, formerly Zhipu AI, a Beijing lab spun out of Tsinghua University. The name comes from the 2021 GLM architecture paper, which trained language models with 'autoregressive blank infilling'. The family has grown from the small open ChatGLM models to the frontier-scale, open-weight GLM-5 generation, GLM-5 through GLM-5.3.

Is GLM open source?

Much of it is open weight, and unusually permissively so. GLM-4.5 through GLM-4.7, GLM-5, GLM-5.1, GLM-5.2 and GLM-5.3 Flash are all published on Hugging Face (under zai-org) under the MIT license, so you can download, self-host, fine-tune, and use them commercially at no license cost. The training data is not released, so they are 'open weight' rather than fully 'open source'. The flagship GLM-5.3 ships under Z.ai's own GLM-5.3 License: MIT-style rights plus a security review for model-as-a-service companies with more than $10 billion in annual revenue.

What is the latest GLM model?

GLM-5.3, released on August 14, 2026, is Z.ai's current flagship: the GLM-5.2 base with much more post-training and a 1M-token context, with a faster-serving GLM-5.3 Prime added in September. GLM-5.3 Flash (320B total / 18B active, MIT) is a separate, smaller multimodal model, the one that ran anonymously as the 'Ox Alpha' stealth model. The newest flagship under the plain MIT license is GLM-5.2.

What is the difference between GLM-4.5 and GLM-4.6?

GLM-4.6 is an incremental upgrade over GLM-4.5 at the same rough scale (about 355B total / 32B active). It extends the context window from 128K to 200K, improves real-world coding and agentic tool use, writes better, and uses roughly 15% fewer tokens per task. Both are open-weight MIT models. GLM-4.5 also has a compact sibling, GLM-4.5-Air (106B / 12B), and a vision model, GLM-4.5V.

What is the difference between GLM-5, GLM-5.1, GLM-5.2 and GLM-5.3?

They share one architecture of about 744B parameters with 40B active. GLM-5 (February 2026, 200K context) started the generation; GLM-5.1 (April) retrained it for long-horizon agent work; GLM-5.2 (June) extended the context to 1M tokens and lifted coding sharply; GLM-5.3 (August) adds much more post-training on the GLM-5.2 base and ships under the GLM-5.3 License instead of MIT. Each has its own page here with Z.ai's benchmarks and live prices.

Is there a free GLM API?

Z.ai's own API lists GLM-4.7-Flash and GLM-4.5-Flash as free (as of October 4, 2026); every GLM-5 model is paid per token, but most GLM weights are free to download and self-host. This site's own free GLM-5.3 Flash API (model id ox-alpha) ran until September 30, 2026; it now returns HTTP 410 Gone, and /api has the migration guide.

How do I use GLM models?

Three ways: (1) use Z.ai's own API at api.z.ai or its GLM Coding Plan inside coding tools like Claude Code and Cline; (2) call a third-party host such as OpenRouter, where GLM-5.3 Flash is z-ai/glm-5.3-flash; or (3) self-host the MIT-licensed open weights from Hugging Face (zai-org) with runtimes like vLLM or SGLang. (This site's own free chat and API were retired on September 30, 2026.)

Who makes GLM models?

Z.ai, the international brand of Zhipu AI, a Beijing-based AI lab founded in 2019 as a spin-out of Tsinghua University's KEG lab. The GLM family is its flagship line of models.

All questions answered The full GLM-5.3 Flash dossier

Build with an open GLM model

You have seen the whole family. The next step is to build something real with it, without wiring up any of the infrastructure yourself. For a hands-on start, O-mega.ai, a site from the same team behind Ox Alpha, has a practical guide to building cheap vision agents with GLM-5.3 Flash, the exact model behind this site.

Build something on Founden

Ox Alpha