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

Z.ai has released the code for GLM-5.3-Flash, the model that first showed up last week under the name Ox Alpha. The public rollout matters because OpenRouter had already put up a free hosted version without saying who built it, and that silence set off a wave of guesses across the industry. Now the company has put its name on it, along with the weights on Hugging Face.

GLM-5.3-Flash is not a small tweak to an old system. Z.ai says it uses a mixture-of-experts design with 320 billion parameters, but only 18 billion are active when it answers a prompt. The model can accept up to 1 million tokens of text, images and video, while its responses can run to 131,072 tokens. That is a very large working surface, and the company is clearly betting that efficiency will matter as much as raw scale.

The efficiency story rests on the attention mechanism, the part of the model that decides what to focus on inside a prompt. Z.ai says it cut the hardware burden with sparse attention, which looks only at the most relevant tokens instead of every single one. It then added linear attention, which changes the way memory grows as prompts get longer. Under that setup, doubling prompt size doubles RAM use, rather than making it balloon more aggressively. The company also says the method swaps out the softmax function for a more efficient approach.

Z.ai says the result is a model that costs 10 times less to run than its previous-generation LLM. It also pointed to benchmark results against Claude Opus 4.8, GPT-5.6 Terra and Gemini 3.7 Flash. GLM-5.3-Flash came out on top on GDPval-AA v2 and placed second on AutomationBench. The company trained it on 30 trillion tokens and says its mHC technique helps keep gradients from getting distorted during training. That is the kind of engineering detail that usually stays buried. Here, it is the whole pitch.