Ox Alpha Ships as GLM-5.3-Flash With Open Weights
The mixture-of-experts system carries 320 billion parameters with 18 billion active, and takes inputs of up to one million tokens.
Illustrative image: a wide server hall lit in cool blue and warm amber, with no identifiable brands.
Z.ai has released the model it had been testing under the codename Ox Alpha as GLM-5.3-Flash, with the weights published openly on Hugging Face.
At a glance
- Architecture: mixture of experts, 320 billion total parameters, 18 billion active per request.
- Context: one million tokens of input across text, images and video; up to 131,072 tokens of output.
- Training: 30 trillion tokens, according to the SiliconANGLE report.
- Benchmarks: top score on GDPval-AA v2, second place on AutomationBench.
- Cost: ten times more cost-efficient than its predecessor, which the report does not name.
Z.ai has put the weights of GLM-5.3-Flash into the open. The model surfaced a week earlier under the codename Ox Alpha, and it now sits on Hugging Face as zai-org/GLM-5.3-Flash. Marketplace operator OpenRouter had already been running a free hosted version of it while the codename still applied.
Inside the model
GLM-5.3-Flash is a mixture-of-experts design holding 320 billion parameters, of which 18 billion stay active on any single request. That gap is where the architecture earns its keep: compute per answer tracks the active slice, not the full weight count. Input runs to one million tokens and covers text, images and video, while output extends to 131,072 tokens.
SiliconANGLE points to three engineering choices behind it. A sparse attention scheme reads only the tokens that matter instead of scanning every one; a linear attention path swaps the softmax function for a cheaper routine; and a method labeled mHC tunes gradients during training. The training run covered 30 trillion tokens.
Benchmark standing
The report credits GLM-5.3-Flash with the top result on GDPval-AA v2, an evaluation aimed at knowledge work, and second place on AutomationBench, which scores how well a model completes tasks inside cloud applications. Claude Opus 4.8, GPT-5.6 Terra and Gemini 3.7 Flash appear as the reference points.
The report gives no numeric scores for either benchmark, and it does not say whether the results have been reproduced independently. Treat the rankings as claims worth re-running locally.
Price and access
The economic argument is efficiency per unit of work: the model is described as ten times more cost-efficient than its predecessor. The report does not identify which earlier model that comparison uses, and it lists no per-million-token rate.
What developers can act on today is narrower and more useful. The weights are downloadable, and OpenRouter's hosted endpoint was free at launch, so evaluating the model against a real workload costs time rather than budget.
What remains unverified
Z.ai's own model page returned no readable text when we tried to open it. License terms, published pricing and the full benchmark table therefore could not be checked against the vendor's wording. Every figure above comes from SiliconANGLE's reporting.
FAQ
What is GLM-5.3-Flash?
A mixture-of-experts model from Z.ai, tested before release under the codename Ox Alpha. Its weights are published on Hugging Face.
How many parameters does GLM-5.3-Flash have?
320 billion in total, with 18 billion active per request. It accepts one million tokens of input and produces up to 131,072 tokens.
Where can I try GLM-5.3-Flash?
The weights are on Hugging Face under zai-org/GLM-5.3-Flash, and OpenRouter offered a free hosted version of the model at launch.
