GLM 5.3

Model ID: vllm/release/glm-5-3 Z.AI (Zhipu) Experimental

EU AI Act

Attribute	Value
Classification	GPAI (General-Purpose AI Model)
Risk class	Limited Risk (Article 50)
Transparency obligations	Yes
noris role	Distributor

This model potentially meets the criteria of a GPAI model with systemic risk (Article 51); the assessment is the responsibility of the model developer. The risk class denotes the transparency level during operation (Article 50). noris acts as a distributor, passes on the developer’s information unchanged, and makes it available pursuant to Article 53(2) and (3).

Technical Specifications

Property	Value
Model name	GLM 5.3
Provider	Z.AI (Zhipu)
Model ID	vllm/release/glm-5-3
Parameters	~753B (MoE)
Architecture	MoE with sparse attention (DSA), 256 experts, 8 active per token
Context length	1,048,576 tokens
Modalities	Text (EN, ZH)
Checkpoint	NVFP4, quantized by Inferact
License	Z.AI license (MIT-style, with a clause for MaaS providers above 10 bn USD revenue)

Recommended Use Cases

• Complex coding and agentic engineering (Terminal-Bench 2.1: 88.2%, DeepSWE: 66.9%)
• Long-running agent workflows (Agents’ Last Exam: 28.5%)
• Security analyses and vulnerability discovery (CyberGym: 84.5%, top score among open-weights models)
• Deep reasoning with controllable effort (reasoning_effort: low, high, max)
• Large document analysis (1M context)

Known Limitations

Caution

• Very large model (high latency): due to its size, inference is slower than with smaller models.
• Primarily EN/ZH: performance is optimized mainly for English and Chinese.
• Text only: the model processes text input exclusively.
• Experimental tier (2-week deprecation notice): this model may be deprecated with only two weeks advance notice.

Note

Thinking-mode control: reasoning_effort accepts low, high, and max and falls back to max if not passed. For chat scenarios, pass clear_thinking=true explicitly in the chat template.

Risk Assessment (EU AI Act)

Risk factor	Category	Level
Model size	Deepfake potential	Medium
Open source (Z.AI license)	Misuse potential	Low
Environmental impact	Model size (753B)	Medium-High
Experimental tier	Deprecation notice	Flagged

Training Data & Sustainability

Training data: Publicly available in the technical report (arxiv: 2602.15763).

Sustainability:

Metric	Value
CO₂	High (753B model), specific figures not published
Optimization	Sparse attention (DSA) keeps inference costs low for long contexts

Original Sources

• Base model: https://huggingface.co/zai-org/GLM-5.3-BF16
• NVFP4 checkpoint: https://huggingface.co/Inferact/GLM-5.3-NVFP4
• License: Z.AI license (see repository)
• Technical report: https://arxiv.org/abs/2602.15763

Model card created October 15, 2026. This model card serves transparency under the EU AI Act (Regulation (EU) 2024/1689). The information is based on publicly available information from the model provider.