Z.ai Co. today released the code for GLM-5.3-Flash, a large language model that’s 10 times more cost-efficient than its predecessor.

The algorithm made its original debut last week under the codename Ox Alpha. LLM marketplace operator OpenRouter Inc. launched a free hosted version of Ox Alpha and didn’t disclose its developer, which drew a significant amount of industry attention. Users soon started speculating that Z.ai is the model’s creator.

GLM-5.3-Flash features a mixture of experts architecture with 320 billion parameters. It activates 18 billion parameters to answer prompts. Users requests can include a 1 million tokens worth of text, images and video while GLM-5.3-Flash’s responses contain up to 131,072 tokens.

The model features a different architecture than Z.ai’s earlier LLMs. One of the biggest changes is in its attention mechanism, a component that analyzes user prompts and extracts the most important details. It identifies those key details by breaking down each prompt into tokens and comparing them against each other.

Analyzing every single token in a lengthy prompt requires a significant amount of processing power. GLM-5.3-Flash reduces that hardware overhead with a technique called sparse attention. Instead of analyzing every single token in a prompt to find important details, the model reviews only the most relevant tokens.

Z.ai further reduced the LLM’s hardware footprint using a method called linear attention. Usually, doubling the size of a prompt quadruples the amount of memory that a model’s attention mechanism consumes. When linear attention is enabled, RAM usage only doubles.

One of the main reasons attention mechanisms are so memory-intensive is that they use an algorithm called a softmax function to interpret prompts. It turns numerical values generated by the host LLM into probabilities.  Linear attention, the technology Z.ai implemented in GLM-5.3-Flash, substitutes the softmax function with a more efficient algorithm.

The company says that the model costs 10 times less to run than its previous-generation LLM. Furthermore, it demonstrated strong performance across a set of popular artificial intelligence benchmarks.

Z.ai compared GLM-5.3-Flash against Claude Opus 4.8, GPT-5.6 Terra and Gemini 3.7 Flash. The former model achieved the highest score on GDPval-AA v2, an evaluation that measures LLMs’ ability to perform knowledge work. GLM-5.3-Flash also placed second on a benchmark called AutomationBench. It’s a test that assesses LLMs’ ability to complete tasks in cloud applications.

Z.ai trained GLM-5.3-Flash on a dataset with 30 trillion tokens. It used a technology called mHC to optimize the workflow.

When an LLM completes a training task, it receives a piece of data called a gradient. The data travels through the model’s artificial neuron layers and reconfigures them to improve their performance. The gradient sometimes becomes distorted along the way, which lowers its effectiveness. The mHC technology that Z.ai implemented in GLM-5.3-Flash lowers the risk of such technical issues. 

GLM-5.3-Flash’s weights are available on Hugging Face.

Image: Unsplash