Z.ai releases GLM-5.3-Flash with frontier performance at lower cost

Z.ai has introduced GLM-5.3-Flash, a new multimodal model in its GLM-5 family. The company published the announcement on August 26, 2026, presenting the model as a lower-cost option for coding, agentic tasks and visual work. Z.ai says GLM-5.3-Flash has 320 billion total parameters, with 18 billion active during inference, and is the first natively multimodal model in the GLM-5 series.

The release combines several engineering changes. Z.ai describes a hybrid attention design that uses linear attention for local dependencies and sparse attention to retrieve relevant information from long contexts. It also introduces IndexPool, which compresses four indexer key vectors into one, and uses Manifold-Constrained Hyper-Connections to improve scaling efficiency. According to the company, the model supports contexts of up to one million tokens and reduces attention compute and KV-cache size compared with GLM-5.3 by factors of about 3.0 and 4.4 respectively.

Z.ai reports gains over GLM-5.2 on six coding and agent benchmarks. Examples include 84.3 versus 81.0 on Terminal-Bench 2.1, 63.4 versus 46.2 on DeepSWE v1.1 and 48.8 versus 26.2 on AutomationBench. In Z.ai's own Code Bench, tested through Claude Code 2.1.207, the company reports a maximum-effort score of 29.0 for GLM-5.3-Flash versus 29.5 for Claude Opus 4.8. These are vendor-reported results; performance can change with prompts, tools, settings and evaluation environments.

The model is also designed to interpret images, documents, charts, interfaces and other work artifacts while checking its own output visually. Z.ai says it served early traffic on a large cluster of Chinese AI chips and improved end-to-end serving performance threefold against its initial baseline on the same hardware. That claim is relevant to inference economics, but it is not an independent hardware benchmark.

For users and makers, the release matters because it connects multimodal capability with open deployment options. Z.ai says the weights are publicly available on Hugging Face under the MIT License, with support for SGLang, vLLM and TokenSpeed. GLM-5.3-Flash is available to GLM Coding Plan users and through ZCode. Lower active parameter counts may reduce serving costs, while the one-million-token context and visual feedback could help with large codebases, presentations and business documents. Independent testing of quality, safety, licensing, hardware requirements and real-world cost remains essential before organisations make it a default model.