GLM-5.3 is now open-weight: Z.ai published the full zai-org/GLM-5.3 repository on Hugging Face on August 28, 2026, two weeks after launching it as a closed API model on August 14. You can pull the raw weights and self-host, or skip the hardware entirely and hit the same model through an API in the next few minutes.
Short answer: GLM-5.3 is now open-weight — Z.ai released the full
zai-org/GLM-5.3weights on Hugging Face on August 28, 2026, under a custom "glm-5.3" license. It's a 753-billion-parameter Mixture-of-Experts model, too large for consumer hardware. For most people, the faster path is Z.ai's own API at $1.40/$4.40 per million input/output tokens, or an OpenRouter provider, not self-hosting.
I've been tracking this release since the August 14 launch, including what GLM-5.3 actually costs against Claude Opus 5 and GPT-5.6 Sol. When the Hugging Face countdown page finally flipped to a real, populated repository, I pulled it up and checked what's actually in it, rather than repeating Z.ai's launch post. Here's what's really there and the fastest way to get running, whether or not you own the GPUs to self-host it.
What you'll need
For self-hosting, you need serious hardware: GLM-5.3 is a 753-billion-parameter model, and even the fp8 checkpoint won't fit on a single GPU or a consumer workstation. Multi-GPU server hardware (think 8-way H200 nodes) is the realistic floor, plus a deployment framework — the Hugging Face card lists support for vLLM, SGLang, Transformers, KTransformers, Unsloth, and Ascend NPU builds. If that's not you, the only things you need are a browser and, for the API, a Z.ai account with a prepaid balance or a GLM Coding Plan subscription. An OpenRouter account works too if you'd rather not create a Z.ai account at all. Nobody needs all three paths — pick the one that matches your hardware and budget before you start.
Step-by-step: getting GLM-5.3 running now that it's open-weight
1. Confirm the weights are actually live
Before you plan around self-hosting, check the repo yourself rather than trusting a blog post — Z.ai's own Hugging Face page ran a countdown for two weeks and slipped past its original target before going live. When I checked huggingface.co/zai-org/GLM-5.3 on August 29, it was a real, populated repository with safetensors files in BF16, F8_E4M3, and F32, not an empty countdown page.
2. Decide if self-hosting is realistic for you
In my testing this is the step people skip. GLM-5.3 is a 753-billion-parameter Mixture-of-Experts model with roughly 40B active parameters per token — the same base architecture as GLM-5.2, carried over into 5.3. The fp8 checkpoint runs to roughly 750 GB on disk and the bf16 version close to 1.5 TB, which means an 8-GPU H200 node at minimum. If you don't already have that kind of cluster sitting around, skip straight to step 4.
3. Pull the weights, if you're self-hosting
Install the Hugging Face CLI and download the repo:
“ pip install -U huggingface_hub huggingface-cli download zai-org/GLM-5.3 --local-dir ./glm-5.3 “
Then serve it with vLLM or SGLang per the deployment instructions on the model card. If your hardware can't fit the full 753B checkpoint, look at zai-org/GLM-5.3-Flash instead — it's a separate 320-billion-parameter, 18B-active model that Z.ai shipped a couple of days earlier under a plain MIT license, and it's the first natively multimodal model in the GLM-5 line.
4. Or skip the hardware and call the API instead
This is what I actually used for testing prompts, and it's the realistic option for almost everyone. Create an API key at Z.ai, then call glm-5.3 against one of three OpenAI- and Anthropic-compatible endpoints: https://api.z.ai/api/coding/paas/v4 (OpenAI Chat Completions-style), https://api.z.ai/api/v1 (OpenAI Responses-style), or https://api.z.ai/api/anthropic (Anthropic Messages-style). If your existing tooling already speaks to Claude or GPT models, you can point it at Z.ai's endpoint with almost no code changes.
5. Or route through OpenRouter for provider choice
I also ran the same test prompts through OpenRouter, which lists GLM-5.3 across more than a dozen hosts — DeepInfra, Fireworks, Friendli, io.net, AtlasCloud, Baseten, and Z.ai itself among them. Pricing there ran closer to $1.20/$4.00 per million input/output tokens when I checked, slightly under Z.ai's own direct rate, because you're comparing across competing providers instead of one fixed price.
6. Check the context window before you build around it
GLM-5.3 supports a 1-million-token context window with a 128K-token maximum output, per Z.ai's own docs. In my testing, that's wide enough for whole small repos in one call, but I still hit the 128K output ceiling once on a long refactor request — worth knowing before you assume a single response can hold everything you asked for.
Example prompts you can copy
These are close to what I ran against the API while testing the steps above:
- Quick capability check: "Summarize what changed between GLM-5.2 and GLM-5.3 based only on the differences you were trained on, and flag anything you're not sure about."
- Coding task: "Here's a failing test and the file it covers: [paste]. Find the root cause before proposing a fix, and don't rewrite anything outside the function that's broken."
- Long-context test: "Here's an entire small repo: [paste multiple files]. List every place a specific config value is read, not just where it's defined."
- Self-host sanity check: "Return a JSON object matching this schema, with no surrounding prose: [schema]." — a good first call to confirm your local vLLM/SGLang endpoint is actually returning clean output.
- Compatibility check: "You're being accessed through an Anthropic-Messages-compatible endpoint. Confirm you can follow a system prompt written for Claude without extra instructions."
Common mistakes to avoid
The first mistake is assuming "open-weight" means "runs on your laptop." It doesn't — 753 billion parameters puts this firmly in server-cluster territory, and I've seen people budget a single high-end GPU for it before checking the file sizes. Second, don't trust a countdown timer as a release date; Z.ai's own Hugging Face page slipped past its original mark, so I'd always confirm the repo has real files before writing anything that depends on a specific date. Third, if you only need to try the model rather than fine-tune or self-host it, don't reach for the weights at all — the API is faster to set up and cheaper for casual use than provisioning GPU hardware you'll barely keep busy. Fourth, watch which license applies to which model: the flagship GLM-5.3 repo is listed under a custom "glm-5.3" license, not MIT, while the smaller GLM-5.3-Flash variant is plain MIT — check the actual repo before assuming the terms match the last model you used. Fifth, if you're price-comparing, use Z.ai's own current pricing page or OpenRouter's live listing, not a launch-day announcement — I found the two differ by about 15% depending on which provider is serving the request.
How to get GLM-5.3 running, compared
| Route | Setup time | Cost | Best for |
|---|---|---|---|
| Z.ai API (direct) | Minutes | $1.40 / $4.40 per 1M input/output tokens | Fastest path, official pricing and support |
| OpenRouter | Minutes | ~$1.20 / $4.00 per 1M tokens (varies by provider) | Comparing providers, avoiding a Z.ai account |
| GLM Coding Plan | Minutes | Points-based quota, 50% off nights/weekends | Heavy coding use inside an existing plan |
| Self-host (Hugging Face weights) | Hours to days | Free license, but ~8x H200 GPUs minimum | Fine-tuning, data residency, no outbound API calls |
Pricing confirmed on Z.ai’s own API pricing docs and OpenRouter’s GLM-5.3 listing, both checked August 29, 2026. The weights themselves are free to download once you accept the license; what you're really paying for with self-hosting is the GPU time, not the model.
Tools that make this easier
If you're weighing GLM-5.3 against other open-weight releases before committing GPU budget or API spend, my breakdown of GLM-5.3’s real cost against Claude Opus 5 and GPT-5.6 Sol covers the benchmark side I didn't repeat here, and DeepSeek V4 Flash’s pricing and benchmarks runs the same kind of independent check on a different open-weight model. For wiring an open-weight API into a coding tool you already use, how to use DeepSeek for coding covers the base-URL swap that works the same way for Z.ai's endpoints. If self-hosting is genuinely on the table, running a local model on consumer-grade hardware is a useful gut check on what "open-weight" costs once you're the one buying compute, and DeepSeek Harness is worth a look if you want an open agent runtime to sit on top of whichever model you land on. For the full field ranked side by side, my running list of the best AI models and Qwen vs. DeepSeek cover where GLM-5.3 fits next to its closest open-weight competitors.
Frequently Asked Questions
GLM-5.3 is now open-weight — is it free?
The weights themselves are free to download once you accept Hugging Face's license terms. Running them isn't free in practice, though — you're paying for GPU time instead of an API bill, and at 753 billion parameters that's real server hardware, not a laptop. The API alternative costs $1.40/$4.40 per million input/output tokens through Z.ai directly.
How long does it take to get GLM-5.3 running?
Through the API, minutes — create a key, pick one of the three compatible endpoints, and start sending requests. Self-hosting is a different scale of project: downloading a 750 GB-plus checkpoint alone takes a while on most connections, and that's before you've set up vLLM or SGLang and confirmed your GPU cluster can actually serve it.
What is the easiest way to try GLM-5.3?
Use the API. Z.ai's glm-5.3 model works over OpenAI-compatible or Anthropic-compatible endpoints, so if you already have a chat client or coding tool pointed at Claude or GPT, changing the base URL and key is usually the whole setup.
Do I need my own GPUs to use GLM-5.3?
No, not unless you specifically want to self-host, fine-tune, or run it with no outbound network access. For everything else, the Z.ai API or an OpenRouter-listed provider gets you the same model without buying or renting an 8-GPU node.
Is GLM-5.3 the same model whether I self-host or use the API?
Yes — the API serves the same weights now published on Hugging Face, not a separate hosted-only version. The one difference to check is which license applies: the full GLM-5.3 repo carries a custom "glm-5.3" license, while the smaller GLM-5.3-Flash variant released alongside it is plain MIT.