Overview
- Tool name: ChatGLM
- Developer: Zhipu AI (Beijing Knowledge Atlas Technology Co., Ltd.)
- Official website: https://chatglm.cn/
- Category: AI Chatbots
ChatGLM, sold in China as Zhipu Qingyan, is the consumer app of Zhipu AI, the Tsinghua spin-off that became the first foundation-model company to list on the Hong Kong Stock Exchange in January 2026. It launched in August 2023 as a bilingual chatbot and has since been rebuilt around an agent loop: the home screen is no longer a chat box but a row of task modes, and the pitch is that a single sentence produces a deliverable.

That pitch is literal. Ask for a research report and it runs a live web search, groups what it finds by claim, and writes a document with citations. Ask for a deck and it drafts an outline first, then renders slides, and lets you edit the outline before anything is drawn. Ask for a data analysis and it reads your spreadsheet, picks a method, and returns charts. The same row of modes covers posters, previewable web pages and document reading, and a separate app, Qingying, handles text-to-video and image-to-video.
The model underneath is GLM-5, released in February 2026: 744 billion total parameters with 40 billion active per token, a 200,000-token context window, and weights published under the MIT license. You do not choose a model in the consumer app, you choose how hard it should think: fast, thinking, deep thinking or deliberation, each with a different cost in time and credits.
The commercial model is a free tier plus two paid levels. VIP is ¥79 a month on auto-renew and includes 10,000 credits a month; SVIP is ¥229 a month with 33,000 credits. Every task that runs long or generates media draws on those credits, which is the part most reviews skip and the part that decides whether the subscription is enough for you.
Key Uses
- Research reports: Research mode searches the live web, clusters sources and writes a structured report with citations you can open one at a time. It is the most reliable of the task modes because the output is a document with visible provenance rather than a chat answer.
- PPT decks: PPT mode takes a one-line brief, proposes an outline, and renders slides only after you approve or edit that outline. Free accounts get a small daily download allowance per file; paid tiers get a much larger one, and re-downloading the same file costs nothing. This ordering - outline before layout - is what keeps a re-run from costing you a whole deck.
- Data analysis: Upload an xlsx or csv, confirm how it read the columns, and it returns cleaning steps, a suggested method, charts and a plain-language read of what moved. The output is a starting point for a human analyst, not an audit.
- Web app prototypes: Webpage mode turns a description into a previewable page you can open in the same window. It is a prototype tool: good for showing a layout to a stakeholder, not a deployment pipeline.
- Posters and long images: A single sentence produces a poster or a long image, with the copy and the layout generated together. Useful for social posts and internal announcements where you need something on brand and today.
- Document reading: Document mode accepts office files, scanned images, ebooks and audio up to 100MB per file, and answers questions with the page in view. In ordinary chat the limit is 10 files at 200MB each, covering pdf, doc, xlsx, ppt, txt and mp4.
- Short video with Qingying: Qingying, Zhipu's video generator, offers text-to-video 2.0 and image-to-video 2.0, with controls for visual style, emotional tone and camera movement, background music, and a paid speed-up queue. Members can extend the acceleration window.
- AI drawing: Members get 2048x2048 high-resolution redraw and local inpainting on desktop; the free tier caps redraw at 1024x1024. It is a utility for fixing an image rather than a full illustration studio.
- Custom agents and scheduled tasks: You can define an agent with a persona, skills and memory, publish it as a shareable card, and attach a schedule so it runs at a fixed time and pushes the result to you. Unread results cancel themselves after a few runs, which is a quiet hint about how much push notification the team expects.
- Everyday chat, writing and code: General Q&A, bilingual drafting, rewriting, translation, and code generation with a sandbox that can run Python. This is the baseline the task modes sit on top of.
Pricing
- Free tier: ¥0. Full chat, web search, file upload and every task mode are reachable, with daily quotas and a credit-based ceiling on the heavy ones, 100MB of cloud knowledge base, redraw capped at 1024x1024, and a daily cap on PPT downloads.
- VIP: ¥79 per month on auto-renew, or ¥119 for a single month, ¥219 for a quarter, ¥859 for a year (about ¥71.6 a month). Includes 10,000 credits a month, 1GB of cloud knowledge base, 2048x2048 redraw, local inpainting, peak-hour priority, a much larger PPT download allowance and video acceleration.
- SVIP: ¥229 per month on auto-renew, and the same ¥229 for a single month. Reported quarterly and annual prices are about ¥649 and ¥2,399 (about ¥200 a month). Includes 33,000 credits a month plus everything in VIP.
- Credit top-ups: ¥29 and ¥109 packs, for people whose usage spikes between billing cycles rather than every month.
- First-purchase pricing: the checkout shows a discounted first month, first three months or first twelve months depending on the term you pick, then reverts to the standard rate.
- Payment: WeChat Pay, Alipay or Apple in-app purchase. Auto-renew charges 24 hours before the period ends; a single-month purchase does not renew.
Prices are the official App Store in-app purchase list for version 3.8.6, checked on 27 October 2026. The web checkout quotes RMB but only after sign-in, so the App Store list is the public price source; the two use the same product names and periods. Two things to know before you budget. First, credits are issued monthly and an in-app member notice states the current allowance explicitly (10,000 for VIP, 33,000 for SVIP, raised from 8,750 and 26,250); long agent runs and video generations are metered against them, so the subscription is a credit allowance, not an unlimited plan. Second, the membership is a consumer subscription and is separate from API billing on bigmodel.cn - a paid chat account gives you no API quota, and an API balance gives you no membership.
Who It Is For
- Students and researchers: The research mode plus document reading covers the boring 80% of a literature review: collecting sources, grouping them, and producing a first draft with citations you can verify. Zhipu's own positioning leans on academic writing, and the four named agent scenarios include academic papers.
- Chinese-language knowledge workers: If the deliverable is a Chinese report, a Chinese deck or a Chinese data readout, this is one of the few assistants where the templates, the writing style and the citation habits are tuned for that audience rather than translated into it.
- Content, ecommerce and social teams: Poster mode and Qingying cover the daily grind of campaign images and short clips without opening a design tool, and the outputs are sized for the platforms the team already posts to.
- Developers prototyping an interface: Webpage mode produces a clickable-looking page in one turn. It is faster than scaffolding by hand when the goal is to show a flow, not to ship it.
- Teams that need a Chinese vendor: Zhipu is a listed Chinese company with domestic data handling and invoicing, which matters for procurement in organizations that cannot buy from an overseas vendor.
- Not for overseas users who need English: The China app expects a mainland phone number or WeChat login and prices in RMB. Zhipu runs a separate international product at z.ai with its own plans; the two do not share a subscription.
- Not for teams that need an API budget: Chat membership and API access are separate products. If your plan is to call a model from code, price it on bigmodel.cn instead of buying VIP seats.
How to Use ChatGLM
1. Open chatglm.cn and sign in with a mainland phone number or WeChat. A new number is registered on first sign-in.
2. Ignore the empty chat box for a moment and look at the mode row above it: Task (the old Agent mode), Research report, PPT, Data analysis, Creative poster, Web app and Document reading. Picking the mode first is the single biggest quality difference, because each one has its own pipeline and its own output format.
3. Write the brief as a deliverable, not a topic. Instead of asking about the industry, ask for a 12-page Chinese report for an internal investment committee, with three competitor profiles and a table of unit economics. Name the audience, the length and the format.
4. Attach the source material. Drag files into the chat (up to 10 files, 200MB each) or put them in the cloud knowledge base first so later conversations can reach them without re-uploading.
5. Choose how hard the model should think. Fast is for lookup and rewriting, Thinking for multi-step work, Deep thinking for analysis and code, Deliberation for the long agent runs. The harder modes take longer and, on metered tasks, cost more credits.
6. For a research report, set the scope before it starts: time window, source types, and whether you want citations inline. Then check the citations by opening them - the report is only as good as the links it used.
7. For a deck, edit the outline before rendering. Change section order, cut a section, add your own numbers; the slides are generated from that outline, so this is the cheap moment to fix structure.
8. For data analysis, upload the sheet and confirm the column types the tool inferred before asking for conclusions. A misread date column produces confident nonsense.
9. For video, write the shot rather than the topic - subject, action, setting, camera move - then pick style, emotional tone and camera movement, generate a short test, and only extend the acceleration window once the look is right.
10. Save what worked. Turn a repeated brief into a custom agent with its own persona and skills, or attach a schedule so the same task runs every morning and pushes the result to you.
Tips for Best Results
- Ask for the artifact, not the answer. 'Draft the slide titles and speaker notes for a 10-page quarterly review' gets a usable file; 'tell me about quarterly reviews' gets an essay.
- Use the outline as a checkpoint. In PPT mode and in research mode, the outline is where you spend five minutes to save a full regeneration.
- Put standing material in the cloud knowledge base. Re-uploading the same PDF in every conversation is the most common waste of the free tier.
- Open the citations. Research mode is good at structure and honest about sources, but it will happily cite a page that says the opposite of the sentence it supports.
- Watch the credit meter on agent and video tasks. A long Task run or a batch of clips can eat a visible slice of the monthly allowance; ordinary chat barely moves it.
- Generate one test shot before committing to a video batch. Style and camera controls change the result more than the prompt does, and re-rolls are billed.
- Use the free tier for a week before paying. It is genuinely usable at low volume, and it tells you whether your work is chat-shaped (stay free) or artifact-shaped (VIP earns its keep).
- Do not buy a year on day one. Buy a month, measure your credit burn against the 10,000 allowance, then decide whether you need SVIP for volume or VIP is enough.
Limitations
- Chinese-first product. The app interface, task templates, prompt library, help center and support run in Chinese, and the models are tuned on Chinese material; adding ?lang=en to a shared or public page switches that page to English, but the workspace itself stays Chinese. It writes English well, yet none of the task pipelines are tuned for an English deliverable.
- Mainland account required. Sign-in is a mainland phone number or WeChat, checkout is in RMB through WeChat Pay or Alipay, and the app is distributed in Chinese app stores. Overseas users are pushed to z.ai, which is a separate product with separate billing.
- Membership is not API access. The subscription buys credits inside the app; programmatic use is billed separately on bigmodel.cn by token. Teams that assume one covers the other are surprised at invoice time.
- Credits are a monthly allowance, not unlimited usage. VIP's 10,000 and SVIP's 33,000 are consumed by agent runs, video and heavy image work, and the allowance is set by the vendor and has already been changed once.
- No refunds. The membership agreement states the service is a virtual good that cannot be returned or exchanged once activated, with refunds only for duplicate charges or a technical fault Zhipu cannot fix.
- Auto-renew is on by default for the discounted terms. It charges 24 hours before the period ends, and canceling stops the next charge but does not refund the current one.
- Hard caps you will hit. 100MB per file in document reading, 10 files and 200MB each in chat, 100MB of knowledge base on the free tier, a daily PPT download limit, and a queue on video generation that paid acceleration only shortens.
- Generated content needs checking. The footer says it plainly: content is AI-generated, verify before use. Reports, financial summaries and data conclusions are drafts, and the citations are a starting point rather than evidence.
- China-specific content rules. Outputs pass through platform review, and the acceptable-use terms are written for mainland regulation, which constrains some research and creative briefs.
Frequently Asked Questions
Is ChatGLM free?
Yes, and the free tier is not a trial. You get the chat, web search, file upload and all the task modes at limited daily quotas, 100MB of cloud storage and 1024x1024 image redraw. The limits you feel first are the daily PPT download cap and the credit cost of long agent runs.
What is the difference between VIP and SVIP?
Credits, mostly. VIP is ¥79 a month with 10,000 credits; SVIP is ¥229 a month with 33,000, and both carry the same feature set - 1GB knowledge base, 2048x2048 redraw, inpainting, peak-hour priority and a much larger PPT download allowance. Buy SVIP only if you are burning through 10,000 credits a month on agent runs or video.
How far do 10,000 credits actually go?
Ordinary chat is cheap; agent runs and video are not. A long Task run or a batch of clips can consume a visible share of the monthly allowance in an afternoon, while a month of Q&A and document reading barely registers. The practical approach is to buy one month, watch the usage page, and only then choose a tier.
How does it compare with Doubao or Kimi on price?
It sits in the middle. Doubao's paid tiers are ¥68, ¥200 and ¥500 a month, Kimi's are ¥49, ¥99, ¥199 and ¥699, and ChatGLM is ¥79 and ¥229. The more useful comparison is what the credits buy: ChatGLM's allowance is explicitly metered against agent and video work, so a cheap tier is not cheap if your work is video.
Can I use it outside China?
Technically you can open the site, but you need a mainland phone number or WeChat to register and RMB payment to subscribe. Zhipu's international product is z.ai, with GLM-5 on a separate plan; a ChatGLM membership does not carry over.
Does a membership include API calls?
No. The membership is an in-app subscription; API usage is billed on bigmodel.cn per token, with its own pricing and its own free tier. The membership agreement says the two are independent, and overlapping benefits are not refunded.
Can I get a refund or an invoice?
Invoices are available through the order page for paid orders. Refunds are narrow: the agreement allows them for duplicate charges or a technical fault that cannot be fixed, and states that activated virtual goods are otherwise non-refundable. Cancel auto-renew before the 24-hour window if you do not want the next period.
Which tasks should I not use it for?
Anything where the citation has to be right without you checking it, anything that needs an English-language deliverable, and anything that has to run from code on a schedule outside the app. It is strong at producing a first draft with structure; it is not an auditor and not an API.

