GLM-5.3 NVFP4

Developers can run it locally for coding, agentic tasks, and other text-based model workloads.

Availability

Model files and local deployment instructions are available on Hugging Face.

Pricing model

Governed by NVIDIA Open Model Agreement

Deployment

Supports local deployment using SGLang, vLLM, Transformers, and other frameworks.

Privacy

Ready for commercial or non-commercial use

Licence

NVIDIA Open Model Agreement (plus original GLM-5.3 license included)

• The announcement says deployment on Ascend NPU is supported by inference frameworks including vLLM-Ascend, xLLM, and SGLang. ^[1]

  - For deployment on the `Ascend NPU` platform, inference frameworks such as vLLM-Ascend, xLLM and SGLang are supported — see [here](https://github.com/zai-org/GLM-5/blob/main/example/ascend.md).

• The announcement provides instructions for serving GLM-5.3 locally with SGLang, vLLM, TokenSpeed, Transformers, KTransformers, and Unsloth.
• GLM-5.3 scores 84.5 on CyberGym in the benchmark table. ^[1]

  GLM-5.3 is state of the art on CyberGym for vulnerability discovery, and its gains are largest further up the exploitation chain, where it more than doubles GLM-5.2 on exploitation benchmarks.

• Z.ai reports a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench.
• The model supports context length up to 1M. ^{[2] [4] [6]}
• Quantization reduces disk size and GPU memory requirements by approximately 1.66x. ^[2]

  This optimization reduces the number of bits per parameter from 8 to 4, reducing the disk size and GPU memory requirements by approximately 1.66x.

• The model is an NVFP4-quantized version of ZAI's GLM-5.3. ^[2]

  The NVIDIA GLM-5.3 NVFP4 model is the quantized version of ZAI's GLM-5.3 model, which is an auto-regressive language model that uses an optimized transformer architecture.

• The model was released on Hugging Face on 09/10/2026. ^[2]

  This model was obtained by quantizing the weights and activations of GLM-5.3 to NVFP4 data type, ready for inference with SGLang.

• The model is ready for inference with SGLang. ^[2]

  This model was obtained by quantizing the weights and activations of GLM-5.3 to NVFP4 data type, ready for inference with SGLang.

• The listed supported hardware microarchitecture is NVIDIA Blackwell. ^[2]

  **Supported Hardware Microarchitecture Compatibility:** <br>

• Z.ai reports a score of 84.5 on CyberGym for GLM-5.3. ^{[3] [5]}
• GLM-5.3 supports controlling its thinking budget with the reasoning_effort parameter. ^{[1] [3]}

  - GLM-5.3 supports controlling the thinking budget through the `reasoning_effort` parameter, which accepts three levels: `low`, `high`, and `max`.

• The announcement lists SGLang, vLLM, TokenSpeed, Transformers, KTransformers, and Unsloth as supported deployment frameworks. ^[3]
• Z.ai reports a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench. ^[3]

  + Stronger Coding: GLM-5.3 is the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on our in-house Z.ai Code Bench.

• GLM-5.3 is presented as an open-weights model, with local serving instructions provided. ^{[1] [3]}

  + Stronger Coding: GLM-5.3 is the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on our in-house Z.ai Code Bench.

• New model releases span quantized language models and image generation Daily brief 2026-09-23