Z.ai open-sources ‘Ox Alpha’ model as GLM-5.3-Flash

Z.ai has released the source code for its new large language model, GLM-5.3-Flash. This model is ten times more cost-effective than its predecessor, Ox Alpha. OpenRouter Inc. provided a free hosted version of Ox Alpha, generating considerable industry interest. Z.ai's GLM-5.3-Flash features a mixture of experts architecture with 320 billion parameters and can process up to 1 million tokens of text, images, or video in user requests, while generating up to 131,072 tokens in its responses.

The model's attention mechanism has been redesigned to reduce hardware overhead by analyzing only the most relevant tokens, using sparse and linear attention techniques. These innovations allow GLM-5.3-Flash to operate at a fraction of the cost and with superior efficiency compared to Z.ai's earlier LLMs. The model has demonstrated strong performance in various AI benchmarks, outperforming competitors such as Claude Opus 4.8, GPT-5.6 Terra, and Gemini 3.7 Flash.

Z.ai trained GLM-5.3-Flash on a dataset of 30 trillion tokens using a technique called mHC to optimize the workflow, which minimizes the risk of gradient distortion during training. The model's weights are now accessible on Hugging Face.