⏱️ Lectura: 8 min
Z.ai released the weights for GLM-5.3 today, the direct update to GLM-5.2, the model that had been the company’s flagship in open weights until now. The release adds to the string of open Chinese models that in 2026 compete head to head with Western labs.
📑 En este artículo
This matters because GLM-5.2 had already shown that a model at that scale could run on consumer hardware thanks to aggressive quantization. GLM-5.3 inherits that philosophy and extends it, according to the announcement published on Z.ai’s official blog.
TL;DR
- Z.ai launched GLM-5.3 on August 14, 2026 as the direct successor to GLM-5.2.
- The model’s weights are publicly available for download and local use.
- It keeps the Mixture-of-Experts (MoE) architecture that has characterized the GLM-5 family.
- It exposes an API compatible with the OpenAI format, which simplifies migrating existing integrations.
- It can be served locally with vLLM or Ollama on Linux, macOS, and Windows.
- GLM-5.2, its predecessor, already ran locally thanks to aggressive quantization.
- Z.ai is the international brand of Zhipu AI, a direct competitor of DeepSeek, Moonshot AI (Kimi), and Alibaba (Qwen).
What happened
The announcement was published on Z.ai’s official blog, where the company presented GLM-5.3 as the next iteration of its flagship line of open models. The company, previously known as Zhipu AI and based in Beijing, has been releasing updates to the GLM series on a quarterly cadence since 2025.
GLM-5.3 arrives just a few weeks after the community finished absorbing GLM-5.2, which already had a strong presence in the open-weight segment. The pace of iteration is itself part of the strategy: Z.ai competes not only on capability but on release frequency against DeepSeek and Moonshot AI.
Background and history
The GLM family began as ChatGLM, an academic project from Tsinghua University that later became Zhipu AI’s commercial product. The company rebranded as Z.ai for its international expansion, separating the Chinese brand (Zhipu) from the global brand aimed at developers outside China.
From GLM-4 to GLM-5, each generation expanded the available context, added agentic capabilities (tool use, code execution, web browsing), and narrowed the gap with closed models on coding benchmarks. Part of the appeal compared to models like Llama or Mistral is the license: the GLM family is usually distributed under permissive terms that ease commercial use without extra legal friction.
The table below summarizes how the two latest versions are positioned:
| Aspect | GLM-5.2 | GLM-5.3 |
|---|---|---|
| Architecture | Mixture-of-Experts (MoE) | Mixture-of-Experts (MoE), same family |
| Distribution | Open weights on Hugging Face | Open weights on Hugging Face |
| API access | Z.ai’s own endpoint | Z.ai’s own endpoint, OpenAI-compatible |
| When to use it | Already validated deployments, stable workloads | New projects that want the latest fixes |
| Limitation | No longer receives the latest improvements | Less proven production adoption time |
GLM-5.3: technical details and performance
GLM-5.3 keeps the Mixture-of-Experts approach: instead of activating all of the model’s parameters on every pass, a router selects a subset of “experts” for each token. This allows models with a very large total parameter count without inference cost growing in the same proportion, which is key for a model at this scale to be viable outside a datacenter.
flowchart TD
A["User request"] --> B["GLM-5.3 router"]
B --> C["Subset of active experts"]
C --> D["Attention layers"]
D --> E["Generated response"]
subgraph MoE
B
C
endTo check how many total and active parameters the download you installed actually has, the exact figure is in the model repository’s config.json: the num_experts, num_experts_per_tok, and hidden_size fields define the router’s real shape. There’s no need to trust marketing numbers: that file is the source of truth, and the community usually publishes GGUF or AWQ quantizations derived from the same checkpoint within days of the release.
💡 Tip: run python -c "import json;print(json.load(open('config.json')))" inside the downloaded model folder to inspect the router’s actual configuration before deploying it.
On benchmarks: Z.ai usually publishes its own SWE-bench, LiveCodeBench, and similar numbers in the release blog post. Since those results vary depending on the harness and benchmark version used, the most reliable way to compare GLM-5.3 against other open models is to run the benchmark that matters to you yourself (for example SWE-bench) against your own use case, rather than taking an isolated number from an announcement.
Getting started
Like the rest of the GLM family, GLM-5.3’s weights are downloaded from Hugging Face. The first step is the same on any operating system: install the Hugging Face client and pull down the model repository.
# Linux / macOS
pip install -U "huggingface_hub[cli]"
huggingface-cli download zai-org/GLM-5.3 --local-dir ./glm-5.3
# Windows (PowerShell)
pip install -U "huggingface_hub[cli]"
huggingface-cli download zai-org/GLM-5.3 --local-dir .\glm-5.3With the weights on disk, the simplest way to stand up an OpenAI API-compatible endpoint is with vLLM. The VLLM_API_KEY variable sets the key you’ll require from clients, and --served-model-name is the name you’ll later use in the model field of each request:
export VLLM_API_KEY="your-local-key"
vllm serve zai-org/GLM-5.3 \
--tensor-parallel-size 8 \
--served-model-name glm-5.3 \
--port 8000Once it’s up, any client that speaks the OpenAI protocol can point to that endpoint by changing only the base URL:
curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer your-local-key" \
-d '{
"model": "glm-5.3",
"messages": [{"role": "user", "content": "Explain what a MoE router is"}]
}'If you don’t have enough GPUs for the full model, Ollama makes it easy to run quantized versions on a laptop; the command is the same on Windows, macOS, and Linux once the client is installed: ollama run glm-5.3.
Impact and analysis
Every new GLM release puts pressure on the price of closed models. When an open model approaches proprietary options in quality, providers that charge per token lose room to justify high rates against alternatives a team can host on its own. That already happened with DeepSeek V4 and Kimi K3 in previous months.
For development teams in Latin America, the availability of open weights like GLM-5.3 means being able to run inference without depending on a foreign provider or the latency of a remote API, something relevant for applications that handle sensitive data or need full control over the deployed model. Also, since the endpoint is compatible with the OpenAI format, frameworks like LangChain or LlamaIndex work without special adapters: you just need to point the client at your own endpoint.
📌 Note: hosting a model of this size in production involves real GPU cost; before migrating, calculate the break-even point between paying per token in a managed API and maintaining your own infrastructure.
What’s next
It’s common for Z.ai to release lighter variants (designed to run on a single GPU) weeks after the base model launch, following the same pattern DeepSeek and Moonshot AI have already used with their own families. It’s also expected that the community will publish third-party quantizations and LoRA fine-tunes of GLM-5.3 in the first days after release.
The other thing to watch is whether GLM-5.3 shows up soon in independent rankings like Artificial Analysis, where GLM-5.2 already had a presence; that will make it possible to compare the real improvement over the previous generation using a consistent methodology instead of relying only on the numbers the maker itself publishes.
📖 Summary on Telegram: View summary
Try it yourself: run huggingface-cli download zai-org/GLM-5.3 today, or read the original announcement on Z.ai’s blog to see the model in action.
Frequently Asked Questions
What is GLM-5.3?
It’s the latest version of Z.ai’s GLM family, a language model with a Mixture-of-Experts architecture distributed with open weights, the direct successor to GLM-5.2.
Is it free?
The weights are free to download so you can run it yourself, though you need the infrastructure to host it. Z.ai also offers its own API endpoint, which is normally billed per token, just like other open model providers such as DeepSeek or Moonshot AI.
How does it differ from GLM-5.2?
Both share the MoE architecture and the open-weight approach. GLM-5.3 is the most recent iteration and incorporates the fixes and adjustments made after GLM-5.2’s release; the fine details of each change are in the official announcement on Z.ai’s blog.
What hardware do I need to run it?
It depends on the variant and quantization chosen. The full model requires several GPUs with enough combined memory; for local use on a laptop, a more aggressive quantization or a reduced version served with Ollama works better.
Is it compatible with the OpenAI API?
Yes, both the endpoint hosted by Z.ai and a self-hosted deployment with vLLM expose the same /v1/chat/completions format used by the OpenAI API, which makes it easy to migrate existing code by changing only the base URL and model name.
Where can I download it?
From the official repository on Hugging Face, under the zai-org organization.
References
- Z.ai Blog: official announcement of the GLM-5.3 release.
- Hugging Face, zai-org: repository with the published weights of the GLM family.
- GitHub, zai-org: code and tools related to the GLM models.
- vLLM: inference engine used to serve the model with an OpenAI-compatible API.
📱 Do you like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de MARIOLA GROBELSKA en Unsplash
0 Comments