Z.ai Releases GLM-5.3-Flash, Revealing ox-alpha as a Native Multimodal Model

GLM-5.3-Flash is now released

Z.ai released GLM-5.3-Flash on August 26, 2026, describing it as the first natively multimodal model in the GLM-5 series. The company says the model has 320 billion total parameters with 18 billion active parameters per token, combining sparse Mixture-of-Experts scaling with a new efficiency-focused architecture.

The release also resolves an identity that had attracted substantial developer attention: Z.ai says GLM-5.3-Flash was the model anonymously tested as ox-alpha on OpenCode and OpenRouter before launch. That makes this a full public release rather than a rumor or benchmark-only preview.

A hybrid attention architecture targets long-context efficiency

GLM-5.3-Flash combines linear attention and sparse attention. Z.ai says linear attention handles local dependencies through state modeling, while sparse attention retrieves globally relevant context through a lightweight indexer. The architecture also introduces IndexPool, which pools indexer key vectors to reduce memory and latency overhead at context lengths up to one million tokens.

The model additionally uses Manifold-Constrained Hyper-Connections (mHC) and a 30-trillion-token multimodal pretraining corpus. Compared with the earlier GLM-4.5 family, Z.ai says GLM-5.3-Flash cuts the activated parameter count from 32B to 18B and reduces the number of layers from 92 to 45.

Coding, agentic and vision benchmarks are strong — but vendor-reported

Z.ai reports substantial gains over GLM-5.2 on its published evaluation suite. Examples include 63.4 vs. 46.2 on DeepSWE v1.1 and 48.8 vs. 26.2 on AutomationBench v1.0.6. It also reports 84.3 on Terminal Bench 2.1 and 78.4 on Toolathlon Verified.

Those results are useful signals, but they should be treated as vendor-reported benchmarks, not independent proof that GLM-5.3-Flash will outperform another model on every codebase or agent workflow. Developers should validate latency, tool reliability, visual understanding and long-context accuracy on their own workloads.

Visual intelligence is part of the model, not a separate variant

Unlike prior text-first GLM releases that depended on separate vision variants, Z.ai presents GLM-5.3-Flash as a native multimodal model. The company highlights document, chart, screenshot, presentation and interface understanding as practical targets, alongside coding and multi-step agent work.

This matters for coding agents because visual context increasingly appears inside the development loop: browser screenshots, dashboards, rendered interfaces, design files and application state can become inputs to the same model that edits code and calls tools.

Z.ai says production serving is running on Chinese AI chips

Z.ai says it has been serving GLM-5.3-Flash at scale on domestically developed Chinese accelerators, using an SGLang-based serving stack with techniques including W8A8 quantization, mixed cache quantization and disaggregated encode-prefill-decode scheduling. The company reports a 3× improvement in end-to-end serving performance over its initial baseline on the same hardware.

That infrastructure claim is also vendor-reported and should not be read as a universal hardware comparison. The more important signal is architectural: GLM-5.3-Flash was designed alongside its inference stack rather than optimized only after training.

Open weights and deployment support broaden access

Z.ai says the model weights are publicly available and that local deployment currently supports SGLang, vLLM and TokenSpeed, with additional inference frameworks planned. GLM Coding Plan users also receive access, and Z.ai says the Flash tier provides three times the usable quota of GLM-5.3.

For developers who previously encountered ox-alpha, the key update is now clear: the anonymous model has a public identity, documented architecture, open weights and a production release path. The practical next step is workload-specific testing, especially for agentic coding, browser/computer use, multimodal documents and long-context tasks.