How to Try GLM-5.3, the Free AI Model That Found a Bug in Cursor

You’ve probably seen a dozen “new AI model just dropped” headlines this year and learned to scroll past most of them. This one’s worth stopping for, because the story isn’t a benchmark chart — it’s that the model reportedly found a real, working vulnerability in Cursor, a coding tool used by millions of developers, within days of release.

The model is GLM-5.3, from the Chinese lab Z.ai (formerly Zhipu AI). It isn’t locked behind an enterprise sales call — there’s a free chat tier live right now, no credit card required. If you’ve never tried a model outside the usual ChatGPT/Claude/Gemini rotation, this is a solid excuse to spend five minutes with one that’s genuinely different.

What Is GLM-5.3, Really?

Most people assume a model announcement means a brand-new AI trained from scratch. GLM-5.3 isn’t that. Z.ai built it on top of its existing GLM-5.2 base, then ran a heavy round of extra training focused on long, multi-step tasks — the kind of work that spans dozens of back-and-forth steps, not a single question and answer. The company says that alone delivered roughly a 50% jump in coding quality, without touching the underlying model.

The headline-grabbing part is what happened next: Z.ai turned the model loose on real-world code review, and it surfaced a genuine security flaw in Cursor, the AI coding editor. That’s not a synthetic benchmark score — it’s a model doing what a human security researcher gets paid to do. Z.ai has since opened a public Security Disclosure Ledger, a running record of vulnerabilities the model finds, which is also why the full weights aren’t shipping immediately: they’re going through a safety review first.

Under the hood, GLM-5.3 supports a 1-million-token context window, native tool calling, and the Model Context Protocol (MCP) — the same open standard Anthropic’s Claude uses to connect an AI to outside tools and data sources. That combination makes it as much a coding-and-agent model as a chat one.

How Does It Work?

Picture GLM-5.3 less as a chatbot that answers one question at a time and more as a junior engineer who’s been told to actually finish the ticket, not just describe how they’d fix it. Its extra “long-horizon” training means it’s been rewarded for staying coherent and accurate across many steps in a row — reading a file, proposing a change, checking whether that change broke something else, adjusting — rather than just producing a plausible-looking first response.

That’s also roughly how it found the Cursor bug: instead of answering a single “is this code safe?” prompt, it worked through a review process step by step, the way a person auditing a codebase moves file to file, testing assumptions as it goes. The large context window helps here too — it can hold an entire codebase in view at once, rather than reasoning about a change in isolation.

How to Try It Yourself

You don’t need to touch a terminal or download any weights to test GLM-5.3 today — a browser is enough.

Go to glm5.app and create a free account. New sign-ups get free credits with no credit card required.

Once you’re in the chat interface, confirm the model selector is actually set to GLM-5.3 — some free chat surfaces (including Z.ai’s own chat.z.ai at the time of writing) still default to the older GLM-5.2, so don’t assume the newest model is selected by default.

If glm5.app is unavailable or you want an even lower-friction option, try rewind.ai’s GLM-5.3 chat, which offers a no-signup free tier with a daily token allowance.

Give it a real coding task: paste in a function with a subtle bug, ask it to review a small file for security issues, or hand it a multi-step task like “refactor this and explain each change.”

If you use an AI coding assistant like Cursor or Claude Code, look into pointing it at GLM-5.3 through its API once general access opens — at the time of writing, the pay-as-you-go API is listed but not yet fully open to the public.

Free access is genuinely usable for chat and coding review, but treat it as a preview rather than the finished experience: the API is still rolling out broadly, and the downloadable weights are being held back for a safety pass — a change from Z.ai’s usual practice of shipping weights immediately.

Tips to Get Better Results

Give it a project, not a puzzle. GLM-5.3 was specifically trained on long-horizon, multi-step work, so it tends to shine more on “here’s a small app, add this feature and tell me what you changed” than on one-line trivia questions.

Double-check which model you’re actually talking to. Because free chat interfaces sometimes default to the previous version, a lackluster first answer might mean you never reached 5.3 at all. Confirm the model selector before judging it.

Try it specifically on code review. Given how it made news, security and bug-finding tasks are a genuinely good stress test — paste in a function and ask it to look for edge cases or vulnerabilities, not just style issues.

Use the large context window on purpose. If you’re working with a long document or multiple files, paste in more than you normally would rather than summarizing first — that’s the scenario the extra training was built for.

Keep a second model open for comparison. Running the same prompt through your usual chatbot alongside GLM-5.3 is the fastest way to see where each is actually stronger.

Closing Thought

You don’t need to switch your default AI tool to get value here — just one real task and five free minutes. Create a free glm5.app account, confirm you’re actually on GLM-5.3, and hand it something with real stakes: a bug you haven’t found yet, or code you’re not fully sure is safe. Comparing what it turns up against what you already knew beats any benchmark chart.