A little mystery, now solved: last week a free model called “Ox Alpha” appeared on OpenRouter with no note on who made it. The thing was fast, it was good, and half the industry spent the week guessing. Now it’s out — Z.ai has released the code, and Ox Alpha is really GLM-5.3-Flash.
The specs are solid. GLM-5.3-Flash is a mixture-of-experts model with 320 billion parameters, of which about 18 billion light up for a single answer. The context window holds up to a million tokens of text, images, and video, and responses run up to 131,072 tokens. The weights? Out in the open on Hugging Face.
The genuinely interesting part sits in the architecture. Z.ai reworked the attention mechanism — the piece that fishes the important details out of a prompt. Two tricks drive the speed. Sparse attention no longer looks at every token, only the relevant ones. And linear attention swaps the compute-hungry softmax function for something leaner. The effect: double the prompt and you only double the memory footprint — instead of quadrupling it, the way it usually works. Net result, the company says the model runs ten times cheaper than its predecessor.
On benchmarks it flexes. Z.ai pitted GLM-5.3-Flash against Claude Opus 4.8, GPT-5.6 Terra, and Gemini 3.7 Flash. On GDPval-AA v2, a test for knowledge work, GLM-5.3-Flash took the top score. On AutomationBench, which measures tasks in cloud applications, it landed second. It was trained on 30 trillion tokens using a technique called mHC that keeps the training process cleaner.
For us the point is less the single number and more the pattern. An open model out of China that beats the pricey Western flagships on individual tasks while costing a fraction — that’s no longer the exception, it’s the rhythm. And with open weights, anyone can host the thing themselves, inspect it, build on it.
Which leaves the question of why a model runs through its test phase anonymously as “Ox Alpha” first. Probably because it lets you gather feedback with no bias — before the name colors the expectation. Clever move. And the curiosity it stirs up is the best advertising you could ask for.
Source: SiliconANGLE: Z.ai open-sources ‘Ox Alpha’ model as GLM-5.3-Flash · GLM-5.3-Flash on Hugging Face