GLM-5.3-FlashOpen Weight

GLM-5.3-Flash, released August 26, 2026, had already been running on OpenRouter as an uncredited stealth model called "Ox Alpha" when Z.ai claimed it as a GLM release earlier that day. At 320B total parameters with 18B active it was a fraction of the size of the 743B GLM-5.3 that preceded it by twelve days, and Z.ai benchmarked it against GLM-5.2 rather than that model: it beat GLM-5.2 on all six tests in the launch chart, most heavily on AutomationBench at 48.8% and DeepSWE v1.1 at 63.4%. Its GDPval-AA v2 rating of 1773 topped the chart outright, ahead of the DeepSeek, Claude, GPT and Gemini entries set beside it.

Elsewhere it sat mid-pack against those closed flagships — 84.3% on Terminal-Bench 2.1, 55.3% on Humanity's Last Exam with tools and 26.3% on Agent's Last Exam were all a few points short of the leaders at the time — and the much larger GLM-5.3 still led it on three of the shared tests. Two things made the release notable anyway. It was natively multimodal with a 1M-token context window and went out under the MIT License with weights on Hugging Face, a pointed contrast with GLM-5.3, which had launched through the GLM Coding Plan and ZCode with no weights at all. And Z.ai said it was trained entirely on Chinese AI chips, the first GLM release the lab made that claim for, which turned an open-weight launch into a data point about how far domestic silicon had come. Access opened the same day across the API, ZCode, chat.z.ai and AutoClaw.