Z.ai revealed that the stealth Ox Alpha preview was GLM 5.3 Flash and released its weights under the MIT license. The vendor describes a natively multimodal mixture-of-experts modelA mixture-of-experts model contains multiple specialized subnetworks and activates a selected subset for each input instead of using every parameter every time. with 320 billion total parametersA model parameter is a learned numerical value that helps determine how an AI model transforms inputs into outputs., 18 billion active per tokenAn active AI model parameter is a model weight selected to participate in computation for a particular token or inference step. and a one-million-token context windowA context window is the maximum amount of tokenized information an AI model can consider during one processing session..
The host's eight-test King Bench retest scored the official API version at 63 out of 80, or 78.75 percent, compared with 70 out of 80 during the stealth preview. Reasoning, math and an end-to-end fine-tuning pipeline remained strong, while several visual and front-end generations were less polished. The host notes that run-to-run variance or serving differences could explain the gap.
For local use, total parameters determine memory capacity while active parameters dominate per-token compute. A four-bit quantizationFour-bit AI model quantization stores model values at roughly four bits of precision to prioritize compact size and efficient inference. may require about 180 GB, while more aggressive dynamic quantizations could approach 100 GB. That puts the model within reach of high-memory unified systems but not ordinary consumer machines.
The video also repeats vendor claims about the scale of the stealth beta, low API pricing and deployment on Chinese AI accelerators. Those claims are not independently verified here. The practical case is narrower: GLM 5.3 Flash offers open weights, low active compute and competitive agentic performance for users who can provide enough memory.
Watch on YouTube
