Z.ai has released the code for GLM-5.3-Flash, an open-source large language model that the company says is 10 times more cost-efficient than its predecessor. The model was originally unveiled last week under the codename Ox Alpha, which drew attention when OpenRouter Inc. launched a free hosted version without disclosing the developer. Users quickly speculated that Z.ai is the creator.
Technical architecture and efficiency
GLM-5.3-Flash features a mixture of experts architecture with 320 billion parameters, but it only activates 18 billion parameters to answer prompts. The model can process input containing up to 1 million tokens worth of text, images, and video while generating responses of up to 131,072 tokens. The technology relies on a different architecture than Z.ai’s earlier models, specifically targeting the attention mechanism.
The attention mechanism analyzes user prompts to extract important details by breaking text into tokens and comparing them. Analyzing every token requires significant processing power, so GLM-5.3-Flash uses a technique called sparse attention to review only the most relevant tokens. The company also implemented linear attention, which reduces memory usage. While standard attention mechanisms can quadruple RAM usage when a prompt doubles in size, linear attention only increases RAM consumption by a factor of two.
This efficiency comes from replacing the standard softmax function with a more efficient algorithm. The company states the model costs 10 times less to run than its previous-generation LLM while maintaining strong performance across popular benchmarks.
Related: Gatik secures $200M to expand driverless fleet
Performance on benchmarks
Z.ai compared GLM-5.3-Flash against Claude Opus 4.8, GPT-5.6 Terra, and Gemini 3.7 Flash. The model achieved the highest score on GDPval-AA v2, an evaluation that measures an LLM’s ability to perform knowledge work. It also placed second on AutomationBench, a test that assesses an LLM’s ability to complete tasks in cloud applications.
The company trained GLM-5.3-Flash on a dataset containing 30 trillion tokens using a technology called mHC to optimize the workflow. This method addresses technical issues that can arise during training, such as gradient distortion. When an LLM completes a training task, the gradient travels through the model’s artificial neuron layers to reconfigure them. Without optimization, this gradient can become distorted, lowering its effectiveness. The mHC technology implemented in GLM-5.3-Flash lowers the risk of such technical issues.
For developers looking to deploy the model, the weights are available on Hugging Face. The open release of GLM-5.3-Flash gives organizations a way to access a large-scale model without the financial burden of running proprietary systems.