What is GLM and how it can help you be more productive
GLM isn’t just “China’s other open-weight model” anymore. What began as bilingual research inside Tsinghua University in 2021 has grown into Z.AI (formerly Zhipu AI): a frontier open-weight lab, the world’s first publicly listed large-language-model company, and a model family that now runs inside third-party coding agents and thousands of self-hosted deployments. On August 14, 2026 Z.AI shipped GLM-5.3 – a post-training-only upgrade that lifted coding scores by half and added cybersecurity skills the company compares to Anthropic’s Mythos line.
In this guide we explain what GLM is in 2026, how the GLM-5.3 and GLM-5.3-Flash models work, what ZCode and the GLM Coding Plan can do, what the plans cost, and how to use GLM well.
What is GLM?
GLM is the family of large language models built by Z.AI, a Beijing company founded in June 2019 by Tsinghua University professors Tang Jie and Li Juanzi as a commercial spin-out of Tsinghua’s Knowledge Engineering Group, led by CEO Zhang Peng. GLM (General Language Model) research began at Tsinghua in 2021 around a bilingual Chinese-English pretraining architecture, at a time when most Western labs treated Chinese as an afterthought. The company traded as Zhipu AI domestically for years before rebranding internationally as Z.AI in July 2025, and it became the world’s first publicly listed large-language-model company on the Hong Kong Stock Exchange in January 2026.
GLM-130B (August 2022) proved the bilingual approach at scale; ChatGLM (2023) turned it into a public chatbot; and since mid-2025 Z.AI has shipped the GLM-4 and GLM-5 generations on a roughly quarterly cadence, alongside CogView/CogVideoX image and video models and the AutoGLM phone-and-computer-use agent line. Unlike OpenAI or Anthropic, Z.AI has a long habit of releasing model weights to the public – which made GLM one of the most-downloaded open model families on Hugging Face.
Today, depending on the plan and surface, GLM can:
- Answer questions & explain concepts
- Write, rewrite, summarize & translate
- Search the live web with citations
- Reason step by step with effort levels
- Process 1M-token codebases & docs
- Write, debug, test & ship code as an agent
- Audit code & reason about vulnerabilities
- Analyze images, video & documents (Flash)
- Operate a phone or browser (AutoGLM)
- Run locally on your own hardware via open weights
GLM-5.3, GLM-5.3-Flash, and the GLM-5 line
GLM-5.3 is unusual among frontier releases: Z.AI reused the exact same base model as GLM-5.2 and put the entire improvement into scaled post-training – more coding environments, longer tasks, stronger verifiers, more reinforcement-learning compute. In the company’s own words: “scaling post-training is all we did for GLM-5.3.” That makes it a controlled experiment in how far a frontier base can be pushed without another pretraining run.
Two notes on the wider lineup. First, GLM-5.3’s weights arrived August 28 under a bespoke GLM-5.3 license – MIT-style terms for almost everyone, except that model-as-a-service operators above $10B in trailing revenue must pass a Z.AI security review first. Second, behind GLM-5 sit GLM-4.7, GLM-4.6, and GLM-4.5/GLM-4.5-Air (GLM-4.7-Flash and GLM-4.5-Air are free or near-free on the API), plus the separate GLM-4.5V/GLM-4.6V vision-language line and the CogView/CogVideoX image and video generators.
How does GLM work?
GLM models are Mixture-of-Experts transformers: GLM-5.3 activates only about 40 billion of its 743 billion parameters per token, which is how a model that size can be served this cheaply. Starting with GLM-5, Z.AI adopted a DeepSeek-style sparse attention mechanism to keep the 1M-token window affordable, and GLM-5.3-Flash goes a step further by borrowing Kimi Delta Attention from Moonshot AI. Reasoning is controlled by an effort setting rather than a separate model line: GLM-5.3 always thinks, at low, high, or max effort.
Cyber capability was the surprise. Z.AI reported that as post-training scaled, security skills grew faster than expected – moving from vulnerability identification to constructing complete exploitation chains – which is why Z.AI delayed open weights by about two weeks after launch for a safety review, unlike GLM-5.2’s same-day MIT weights. A typical request flows through several stages:
- 1
Understand the request
Reads the prompt, conversation history, uploaded files, and any connected tools or MCP servers. With 1M context you can feed an entire codebase.
- 2
Reason through the problem
The effort setting (low / high / max) controls how much the model deliberates before answering – GLM-5.3 can no longer skip thinking entirely.
- 3
Decide which tools to use
Web search, code execution, file analysis, or handing off to AutoGLM for on-device phone or browser actions.
- 4
Generate the response or take the action
The answer, the pull request, the vulnerability report, or the completed multi-app task.
- 5
Refine with follow-ups
Long agentic sessions inside ZCode or third-party coding agents stay coherent across the full 1M-token window.
What can you do with GLM?
The agentic shift: ZCode, the Coding Plan, and open weights
Z.AI’s biggest shift mirrors rivals: from “you ask, it answers” to “you define a goal, it ships code.” ZCode, launched July 2, 2026 as an agentic development environment, is the official harness for GLM-5.3 – and unusually, the same models also power third-party agents, so the ecosystem compounds instead of fragmenting.
The strategy is one flagship model scaled via post-training, one cheap multimodal Flash for visual coding, and one IDE that keeps everything in the same loop – with open weights so enterprises are not locked into a single vendor’s cloud.
GLM timeline: 2019–2026
Tang Jie and Li Juanzi found Zhipu AI in Beijing as a commercial spin-out of Tsinghua University’s Knowledge Engineering Group.
GLM research begins at Tsinghua KEG; GLM-130B (August 2022), a 130-billion-parameter bilingual model, is released open-source – the architecture’s first proof at scale.
ChatGLM launches as a public chatbot, followed by ChatGLM2/3 and CogVLM-17B, Zhipu’s first vision-language model.
GLM-4 extends context to 128K tokens and adds multimodal understanding plus “All Tools” function calling.
The US Commerce Department adds Zhipu AI to its Entity List; the company disputes the decision and keeps shipping – training and serving on Huawei and Cambricon chips.
Zhipu rebrands internationally as Z.AI alongside GLM-4.5 (355B/32B active) – the first GLM with one-click third-party coding-agent compatibility.
GLM-4.6 raises context to 200K; AutoGLM is open-sourced in December as the first stable phone-use agent; GLM-4.7 closes out the year focused on end-to-end coding.
Z.AI lists on the Hong Kong Stock Exchange on January 8 – the world’s first publicly listed large-language-model company.
GLM-5 launches February 11 (744B/40B-active MoE, 28.5T tokens); GLM-5.1 ships, trained entirely on non-Nvidia hardware and closing in on Claude Opus-class coding scores.
GLM-5.2 (June 13) adds a 1M-token context window; ZCode launches July 2 as the agentic IDE – then, in July, gets self-hosted by Hugging Face to help contain an autonomous attack by escaped third-party models.
GLM-5.3 launches August 14 with major coding and cybersecurity gains from post-training alone; GLM-5.3-Flash follows August 26 as the first natively multimodal GLM-5 model; GLM-5.3’s own weights arrive August 28 after a safety review.
How to use GLM well
The prompt pattern that works
Role, task, format – with GLM-5.3’s coding and security strengths, a structured review prompt like this is where it earns its price. For long-running agent work, keep the system prompt stable so caching keeps costs down.
How large is the context window?
GLM-5.3 and GLM-5.3-Flash both carry a standard 1,048,576-token context window – roughly 750,000 words – with up to 128,000 tokens of output. GLM-4.6 and GLM-4.7 support 200,000 tokens (about 150,000 words); GLM-4.5 and the original 2023 ChatGLM topped out at 128,000 and 8,000 tokens respectively.
Practical rule of thumb: a token is roughly 0.75 English words, and about 1.5 tokens per word once you factor in punctuation and spacing. As with any 1M-token model, attention is not perfectly uniform across the full window – expect some “lost in the middle” degradation on very long documents.
What does GLM cost?
| Plan | Price | Model access | Limits |
|---|---|---|---|
| Z.AI chat | $0 | GLM-5.3 chat, web search | Usage caps; no coding-agent tooling |
| Coding Plan – Lite | $18/mo | GLM-5.3, GLM-5.3-Flash | Points quota; off-peak calls cost roughly half |
| Coding Plan – Pro | $80/mo | GLM-5.3, GLM-5.3-Flash | Larger allowances, priority access |
| Coding Plan – Max | $168/mo | GLM-5.3, GLM-5.3-Flash, peak priority | Highest quota, team features |
| API – GLM-5.3 | $1.40 / $4.40 | per 1M tokens in/out, 1M context | $0.26 cached input · text only |
| API – GLM-5.3-Flash | $0.15 / $0.50 | per 1M tokens in/out, 1M context | $0.03 cached input · multimodal |
| API – GLM-4.7-Flash | Free | 200K context | Free tier, lower priority |
| Self-hosted | Free (license varies) | Full weights, unlimited use | GLM-5.3-Flash is plain MIT; GLM-5.3 needs a security review above $10B revenue |
Off-peak Coding Plan usage deducts roughly half as many credits per call, so weekend and evening agent runs go further. Calls to legacy GLM-5.2, GLM-5.1, and GLM-4.7 on the Coding Plan are automatically routed to GLM-5.3 or Flash. The chart above reflects mid-September 2026 – check Z.AI’s official pricing page for current rates.
Frequently asked questions
Heads-up:GLM ships new versions on a roughly quarterly cadence, and pricing, licensing, and model IDs change accordingly. The details above reflect September 2026 (GLM-5.3 released Aug 14, weights Aug 28). Check Z.AI’s official docs and Hugging Face pages for current terms.