GLM-5.3 open weights land on Hugging Face after Z.ai's two-week wait (2026)

Z.ai's flagship GLM-5.3 is now available as open weights on Hugging Face. Z.ai announced the model on August 14 and said it would publish the weights after two weeks of safety evaluation and hardening. Hugging Face's model listing records the main repository as updated at 14:48 on August 28, completing that release window. This is the full GLM-5.3 release, separate from the smaller GLM-5.3-Flash edition covered earlier. 1 2

What launched

Signal	Confirmed detail	Why it matters
Weights and size	The main repository contains FP8 weights, while `GLM-5.3-BF16` provides BF16 weights. The Hugging Face model card lists 753B parameters; Z.ai's official repository labels both versions `744B-A40B`. 2 3	The two first-party pages report different total-parameter figures. Treat the model as flagship-scale and verify the exact serving footprint before planning a local deployment.
What changed	GLM-5.3 keeps the GLM-5.2 base model. Z.ai says every gain comes from post-training, with stronger complex coding and long-horizon agent work. 1	This is a post-training update rather than a new base architecture, so GLM-5.2 evaluation suites are a useful starting point for comparison.
Runtime behavior	The API accepts text-only inputs, supports a 1M-token context and 128K maximum output, and keeps reasoning enabled. `reasoning_effort` offers `low`, `high`, and `max`; applications that sent `thinking.type: "disabled"` must migrate. 4	Long context and adjustable effort support long-running coding tasks, while always-on reasoning changes latency and integration assumptions.
Access	The model ID is `glm-5.3` on Z.ai's API. Z.ai and the model card list local serving through SGLang, vLLM, TokenSpeed, Transformers, KTransformers, and Unsloth, with additional Ascend support. 2 4	Developers can test the same release through an API or locally, but the model's size makes framework and hardware compatibility an immediate gate.

Z.ai reports a 50% improvement over GLM-5.2 on its private Code Bench. Its public comparisons report GLM-5.3 at 28.3 on Terminal-Bench 3.0, 66.9 on DeepSWE v1.1, and 28.5 on Agents' Last Exam. Z.ai also reports 84.5% on CyberGym versus 77.2% for GLM-5.2, and 54.4% on ExploitBench versus 24.4%. These are lab-reported results under benchmark-specific protocols, not an independent audit. 1

Z.ai's six-panel benchmark comparison for GLM-5.3 — *Z.ai's chart compares GLM-5.3 with GLM-5.2 and other models across coding, agent, and cyber-related tasks; the scores are Z.ai's reported comparisons. 3*

Why this matters

The open-weights frontier now has a flagship-scale model whose advertised gains target the work developers increasingly hand to agents: multi-step coding, terminal operations, and vulnerability research. The next useful step is a controlled hands-on comparison with GLM-5.2: run the same coding tasks at each reasoning level, measure output tokens and latency, then test the 1M-token path and the serving stack you can actually operate. Cyber capability deserves a separate safety review before the model is exposed to production repositories or live infrastructure. 1 4