Mysterious Ox Alpha AI Revealed as GLM-5.3-Flash: 320B Open Model Challenges Claude

Z.ai has officially unveiled GLM-5.3-Flash, an open multimodal model, revealing the identity behind the mysterious Ox Alpha AI that underwent blind testing on OpenRouter. Featuring a Mixture-of-Experts architecture and 320 billion parameters, the model approaches the coding capabilities of top proprietary flagships like Claude Opus 4.8 while consuming ten times fewer compute resources.

With the free release of its weights, Ox Alpha’s anonymous testing phase has officially ended.

MoE Architecture and Hybrid Attention Trained on 30 Trillion Tokens

The model was trained on a multimodal dataset of 30 trillion tokens, gaining native support for text and image recognition.

Thanks to its Mixture-of-Experts (MoE) architecture, only 18 billion of its total 320 billion parameters are activated per token, keeping hardware requirements for inference relatively low.

To reduce the cost of processing long context windows of up to 1 million tokens, the developers combined sparse (sparse) and linear (linear) attention for the first time. The source code and weights are available on Hugging Face under a permissive MIT license that permits commercial use.

Programming Benchmark Results

In terminal command execution tests on Terminal-Bench 2.1, the model scored 84.3 points, while achieving a 63.4% success rate in the comprehensive software engineering benchmark DeepSWE.

According to Z.ai, the model rivals top commercial AIs in real-world coding tasks, while token generation costs are 10 times lower compared to GLM-5.2.

API Availability and Local Deployment

Following the announcement, the model is available under its official name in the OpenRouter service catalog and through Z.ai’s own platform.

For local deployment and integration into enterprise environments, support for the vLLM, SGLang, KTransformers, and TokenSpeed libraries is provided.