On August 28, Zhipu will release GLM-5.3's full model weights on Hugging Face. The API has been live for over a week. The anonymous Ox Alpha test—confirmed as GLM-5.3-Flash—ran for five days, processing over 50 trillion tokens and attracting millions of users. On August 26, Zhipu confirmed its identity. On August 28, the weights arrive.
This is not a model launch. It is a four-phase rollout: release, API, anonymous validation, open-source drop. Each phase served a purpose.
60 on AA Index, 84.5% on CyberGym—The Numbers Are Real
GLM-5.3 runs on the same 743-billion-parameter MoE base as GLM-5.2, activating roughly 40 billion parameters per token. The improvements come entirely from post-training scaling—longer training runs, more diverse environments, and refined reinforcement learning.
The model's 60-point AA score puts it in the same range as Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol. On Terminal-Bench 3.0, it jumped from 4.6 to 28.3. On DeepSWE v1.1, it moved from 46.2 to 66.9. Internal coding tests show a 50% improvement over GLM-5.2.
In cybersecurity, GLM-5.3 scored 84.5% on CyberGym, slightly ahead of Anthropic's Mythos 5 (83.8%) and OpenAI's GPT-5.6 Sol (83.6%). The model identified over 2,400 vulnerabilities across 269 open-source projects, including a 45-year-old flaw in the Linux kernel.
The API Was the Trial. The Weights Are the Ownership.
The open release is the final step in Zhipu's playbook. The API has been live. Developers have tested the model. The anonymous Ox Alpha run has already validated its performance at global scale. Now the full weights become available for any developer to download, modify, and deploy.
Zhipu has positioned this release strategically. The model is not just competitive with frontier U.S. models—it is priced at roughly 1/40 of Claude Opus 4.8. With the weights open, developers can run it on their own infrastructure, eliminating the API cost entirely.
The release also completes a narrative arc. GLM-5.2 was the model Hugging Face turned to for forensic analysis after OpenAI's agent breach. GLM-5.3 is the model that proves Zhipu can scale that capability for global production use.
Zhipu Just Completed the Perfect Launch Sequence
GLM-5.3's weight drop is the final signal in a perfect launch sequence: release, API, anonymous validation, and open-source drop. Developers have already validated its performance. Tomorrow, they will have the keys to build on it.
The model is open. The hardware is domestic. The price is negligible. The question is no longer whether GLM-5.3 can compete—it already has. The question is what developers will build when the gates open.
P.S. The free Ox Alpha preview ended on August 26, and the API pricing has officially taken effect. But tomorrow, the weights arrive. Developers who burned through trillions of tokens for free this week now face a choice: pay for Zhipu's service, or run the model themselves. Zhipu is betting they'll do both—and building an ecosystem around a model that was already battle-tested before its weights were ever released.
Frequently Asked Questions
Q: When will GLM-5.3 weights be released?
A: Zhipu will release GLM-5.3's full model weights on Hugging Face on August 28, 2026.
Q: How does GLM-5.3 perform compared to other models?
A: GLM-5.3 scores 60 on the Artificial Analysis Intelligence Index, tying it with Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol. On CyberGym, it scored 84.5%, slightly ahead of Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%).
Q: What is the architecture of GLM-5.3?
A: GLM-5.3 runs on a 743-billion-parameter MoE base, activating roughly 40 billion parameters per token. The improvements over GLM-5.2 come entirely from post-training scaling, not architectural changes.
Q: What were the key benchmark improvements?
A: Terminal-Bench 3.0 jumped from 4.6 to 28.3. DeepSWE v1.1 moved from 46.2 to 66.9. Internal coding tests show a 50% improvement over GLM-5.2.
Q: What is the Ox Alpha connection?
A: GLM-5.3-Flash was tested anonymously as Ox Alpha for five days before Zhipu confirmed its identity. The model processed over 50 trillion tokens and attracted millions of users during the anonymous test.
Q: What is the significance of the weights release?
A: The weights release means developers can download, modify, and deploy GLM-5.3 on their own infrastructure, eliminating API costs entirely. It completes Zhipu's four-phase launch: release, API, anonymous validation, and open-source drop.
Q: What is the API price?
A: GLM-5.3 is priced at roughly 1/40 of Claude Opus 4.8. With the weights open, developers can run the model on their own infrastructure.
Q: What is the four-phase launch sequence?
A: Release (August 14), API (August 19), anonymous Ox Alpha test (August 20-24), and open-source weight drop (August 28). Each phase built on the previous one.
Q: Why does this matter for the AI industry?
A: Zhipu just completed a perfect launch sequence: release, API, anonymous validation, and open-source drop. The model was already battle-tested at global scale before its weights were ever released. The question is now what developers will build with it.
