GLM-5.3-Flash Revealed as Ox-Alpha
The mysterious Ox-Alpha model has been revealed as GLM-5.3-Flash, a 320B-parameter MoE model with only 18B active parameters, described as ultra-low-cost. Testing involved 6-7 scenarios, including editing a full movie into a 3-minute commentary. An article published the test prompts, production ideas, and tools. The entire deployment ran on Chinese domestic cards, with an estimated requirement of tens of thousands of cards.
Key insights
GLM-5.3-Flash is the Ox-Alpha model, a 320B MoE with 18B active parameters, ultra-low-cost, and fully deployed on Chinese domestic hardware
The model is confirmed to have 320B total parameters and 18B active parameters, using a Mixture of Experts architecture to achieve flagship-level performance at ultra-low cost. Practical tests included 6-7 scenarios, notably editing a full-length film into a 3-minute commentary. The entire deployment was done on Chinese domestic cards, requiring tens of thousands of them, which signals a move toward domestic hardware self-sufficiency for large-scale AI inference.