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

Ox Alpha’s initial debut came through LLM marketplace operator OpenRouter Inc., which launched a free hosted version without disclosing the developer. The anonymous release drew significant industry attention, and users soon speculated that Z.ai was behind it. Z.ai has now confirmed its authorship by publishing the model under the GLM-5.3-Flash name.

GLM-5.3-Flash uses a mixture-of-experts architecture with 320 billion total parameters, activating 18 billion parameters per prompt. It can process up to 1 million tokens of input, including text, images and video, and generate responses of up to 131,072 tokens. The new architecture differs notably from Z.ai’s earlier LLMs, particularly in its attention mechanism, which analyzes user prompts to extract key details.

To reduce computational overhead, the model employs sparse attention, which reviews only the most relevant tokens rather than every token in a prompt. It also uses linear attention, a method that lowers memory usage scaling from quadratic to linear. While traditional attention mechanisms rely on a softmax function to convert numerical values into probabilities, linear attention substitutes this with a more efficient algorithm.

According to Z.ai, GLM-5.3-Flash costs ten times less to run than its previous-generation LLM. The company benchmarked it against Claude Opus 4.8, GPT-5.6 Terra and Gemini 3.7 Flash, and reported that GLM-5.3-Flash achieved the highest score on GDPval-AA v2, an evaluation measuring knowledge-work ability. It also placed second on AutomationBench, which tests LLM performance in cloud applications.

The model was trained on a dataset containing 30 trillion tokens. Z.ai used a technology called mHC to optimize the training workflow, which reduces the risk of gradient distortion during backpropagation. The model’s weights are available on Hugging Face.