What Is GLM 5.3 and Zcode? ZAI's Coding Model Explained
GLM 5.3 is ZAI's post-trained coding model, and Zcode is its dedicated agentic dev app. Here's how they work and what sets them apart.
What is GLM 5.3?
GLM 5.3 is a coding-focused large language model from ZAI, built on the same parameter count and architecture as its predecessor, GLM 5.2. It isn’t a new base model. It’s GLM 5.2 with substantially more post-training applied, aimed specifically at making it better at agentic coding tasks like exploring a codebase, planning changes, writing code, and verifying the result. ZAI has also positioned it as particularly strong at vulnerability auditing, scanning an existing repository and flagging security issues, which is one of the more specialized use cases the model was tuned for.
What is Zcode?
Zcode is ZAI’s own desktop application for agentic development, built specifically to run GLM models. It is not a chat interface bolted onto a website, and it is not a fork of VS Code. It’s a purpose-built environment designed around how GLM actually works, rather than GLM being squeezed into a harness (like many coding agents) that was originally designed around Claude.
That distinction matters more than it sounds. Coding agents built for one model family often carry assumptions, prompt structures, and tool-calling conventions tuned to that model. Running a different model through the same harness can leave capability on the table. Zcode is ZAI’s attempt to close that gap by owning both the model and the environment it runs in.
How does Zcode work?
Zcode ships with a set of features aimed at real, ongoing development work rather than one-off code snippets:
- Project indexing, so the agent has context on the codebase it’s working in.
- Skills and MCP servers, letting you extend what the agent can do and connect it to external tools and data sources.
- Custom commands for repeatable workflows.
- Usage tracking, so you can see token consumption in real time.
- A built-in preview with dev tools, useful for web app development where you want to see changes rendered immediately.
- Git integration and browser automation.
- Goal Mode, a mode built for larger, multi-step objectives.
One coffee. One working app.
You bring the idea. Remy manages the project.
Goal Mode is worth calling out specifically. Instead of responding to a single prompt and stopping, it keeps the agent oriented toward a broader objective: it builds a task list, tracks progress against it, uses a read-only “explore” sub-agent to investigate the codebase before making changes, and keeps working until the goal is complete or it hits a genuine blocker. This is the mode designed for long-horizon work, the kind of task where an agent needs to hold context across many steps rather than execute one instruction and quit.
Zcode also supports remote control from a phone, with bots for messaging platforms including WeChat, QQ, and Telegram, so a long-running task can be kicked off and monitored from somewhere other than the desktop. The app runs on macOS (both Apple Silicon and Intel), 64-bit Windows, and Linux on ARM and Debian-based systems.
How does Zcode differ from Claude Code?
The core difference is fit between model and harness. Claude Code is built around Anthropic’s Claude models, with its own conventions for tool use, context management, and agent behavior. Running a different model through that same harness works, but it isn’t necessarily optimized for that model’s specific strengths or quirks. Zcode exists because ZAI wanted an environment built from the ground up around GLM’s behavior, rather than adapting GLM to fit somebody else’s harness.
Feature-wise, Zcode currently covers a lot of the same ground: project indexing, custom commands, MCP support, git integration. One notable gap is custom, user-defined sub-agents. Zcode gives you the built-in explore sub-agent, which is read-only and useful for investigation, but if you’re looking for a multi-agent orchestration setup where you define and coordinate several specialized custom agents, that capability isn’t there yet.
What did the GLM 5.3 free token promotion actually offer?
ZAI ran a limited-time promotion called the GLM 5.3 times Zcode weekend build, aimed at new users. The mechanics were simple on paper: log into Zcode for the first time during the promotional window, and 100 million free GLM 5.3 tokens landed in the account automatically, no claim process or payment details required.
The first attempt at this promotion ran for roughly an hour before ZAI pulled it, citing demand exceeding expectations and adjustments to service capacity. Anyone who signed up in that window kept the tokens; anyone after got dropped back to the standard trial. A second round followed with a hard cap of 50,000 token packs, distributed first come, first served, running across a roughly two-and-a-half day window.
A few conditions applied across both rounds: the offer was for new users only, tokens could only be spent inside the Zcode app itself (not piped out through the API to other tools), and unused tokens expired when the event window closed. The tokens behaved like a large daily allowance refreshing at a set time each day, not a static balance sitting untouched until used.
Is GLM 5.3 worth trying outside the promotion?
Even without a limited-time token grant, Zcode’s standard new-user trial gives a meaningful amount of runway: several days of daily token allowances split between GLM 5.3 and a lighter GLM 5 Turbo variant. That’s enough to build something real, not just a toy demo, and it’s a reasonable way to evaluate whether GLM 5.3 fits your workflow before committing to a paid plan.
Beyond the trial, ZAI runs a tiered coding plan (light, pro, and max tiers), with periodic promotional discounts and quota boosts. One detail worth knowing if you’re optimizing usage: peak and off-peak pricing exist, with calls made outside a specific peak window costing fewer quota points. For anyone not on the same time zone as that peak window, working during off-peak hours effectively stretches a quota further at no extra cost.
ZAI has also indicated the model’s weights will be released on Hugging Face some weeks after launch, which would open the door to running GLM 5.3 through self-hosting or third-party providers rather than exclusively through Zcode.
What’s the smartest way to use a large token grant like this?
A large token allowance is easy to waste on trivial prompts. The more useful approach is to treat it like a real development sprint: pick an actual project, whether that’s a small CRM, a bug tracker, a dashboard, or an existing codebase you want audited for vulnerabilities, and work through a structured loop of exploring the codebase, planning changes, building, and verifying the result. Since agentic coding tools operate on live code, it’s also worth applying basic judgment about what you feed into a free, promotional endpoint. Personal, open-source, or test projects are a safer fit than proprietary company code or anything involving customer data, particularly without having reviewed the provider’s data handling policy first.
Frequently Asked Questions
What’s the difference between GLM 5.3 and GLM 5.2?
GLM 5.3 uses the same architecture and parameter count as GLM 5.2. The difference is additional post-training focused on improving agentic coding performance, rather than a change to the underlying model.
Can I use the free GLM 5.3 tokens outside of Zcode?
No. The promotional tokens were restricted to use inside the Zcode app itself and could not be routed through the API into other tools like Claude Code or custom scripts.
Do unused promotional tokens carry over after the event ends?
No. Tokens from the promotional grant expired when the event window closed. They functioned as a large daily allowance during the promotion, not a permanent balance.
Is there still a free way to try GLM 5.3 after the promotion ends?
Yes. Zcode’s standard trial for new users includes a multi-day allowance of GLM 5.3 and GLM 5 Turbo tokens, separate from the limited-time promotional grants.
Does Zcode support custom multi-agent workflows?
Not yet. Zcode includes a built-in, read-only explore sub-agent for investigation, but it does not currently support user-defined custom sub-agents or full multi-agent orchestration.
