Z.ai Releases GLM-5.3-Flash With Open Weights
The 320B mixture-of-experts model pairs native multimodality and one-million-token context with aggressively priced hosted inference.
A smaller active footprint
Z.ai released GLM-5.3-Flash, a natively multimodal mixture-of-experts model with 320 billion total parameters and 18 billion active parameters. The company also published its weights and introduced hosted pricing of $0.15 per million input tokens, $0.50 per million output tokens and $0.03 per million cached input tokens.
The model combines linear and sparse attention to reduce the cost of processing long contexts. A mechanism called IndexPool compresses groups of index keys before retrieval, while manifold-constrained hyper-connections are intended to improve information flow and scaling stability. GLM-5.3-Flash supports a context window of up to one million tokens and was pretrained on a 30-trillion-token multimodal corpus.
Coding, vision and domestic infrastructure
Z.ai says the model substantially improves on GLM-5.2 across coding and agent benchmarks. Its reported scores include 63.4 on DeepSWE 1.1, versus 46.2 for GLM-5.2, and 48.8 on AutomationBench, versus 26.2. The company also reports stronger document, chart and visual-interface capabilities, including workflows where the model inspects a rendered interface and iteratively corrects its own code. These remain vendor-reported results, and comparisons may depend on harnesses, reasoning budgets and context-management settings.
Before the public release, the model appeared anonymously as Ox Alpha on OpenCode and OpenRouter. Z.ai says all of that trial traffic was served on Chinese accelerators. Its production system separates multimodal encoding, prompt prefill and decoding into independently scheduled pools across tens of thousands of domestic chips, with the company claiming a threefold improvement over its original serving baseline.
Why it matters
GLM-5.3-Flash links three strategic developments: open model weights, sharply lower agent-inference pricing and credible large-scale service on non-Nvidia hardware. The model’s public weights make its architecture inspectable, while the hosted price could pressure competitors serving coding and computer-use agents, where long trajectories quickly accumulate tokens. Its most important claim is therefore operational rather than a single benchmark result: that a competitive multimodal agent model can be served economically across a very large cluster of Chinese accelerators.