Chinese AI startup Z.ai, also known as Zhipu AI, officially revealed that the mysterious Ox Alpha model is the latest iteration of its GLM series, named GLM-5.3-Flash [1, 2, 3, 4, 5, 6, 7, 8]. The model was anonymously launched on August 20, 2026, on OpenRouter and OpenCode platforms, rapidly attracting developers and users testing its performance [1, 2, 4, 6, 8].
GLM-5.3-Flash supports multimodal input including text, image, and video. It features an exceptionally large context window exceeding one million tokens (1,050,000 tokens), enabling advanced reasoning over vast data chunks [1, 4, 6, 7, 8]. During its roughly one-week free trial, Ox Alpha became the top AI model on OpenRouter, with usage more than twice that of DeepSeek, a competing Chinese model [1, 2, 4, 5, 6, 8].
Z.ai committed to publicly releasing the trained weights of GLM-5.3-Flash, allowing developers to deploy and build applications on the model. The model files exceed 300GB, requiring substantial hardware resources to run locally [1, 2, 4, 6, 7]. The entire system runs on domestically developed Chinese AI chips linked through proprietary high-bandwidth interconnects [1, 4, 5, 8].
Compared to its predecessor GLM-5.2, GLM-5.3-Flash roughly doubles performance while costing a tenth as much in computational resources and pricing. Token costs are $0.15 for input and $0.50 for output per million tokens, more affordable than comparable U.S. AI models [1, 2, 8]. Artificial Analysis ranked Ox Alpha with a score of 57, aligning it with OpenAI's GPT-5.6 Terra and above Google's Gemini 3.7 Flash [8]. The AI showed notable strength in complex programming and agent tasks, surpassing some major competitors including U.S. counterparts [3, 4, 8].
Before Z.ai’s confirmation on August 26, Ox Alpha’s developer identity was unknown, leading to speculation that it might be Microsoft’s unreleased MAI model. The codename "Ox Alpha" stemmed from a recent popular Chinese animated movie "Niu Lai" (meaning "Ox Coming") [1, 2, 3].
Z.ai, which went public in Hong Kong in January 2026 and rebranded from Knowledge Atlas Technology the previous month, reported 2025 revenue of about 724 million RMB (~$108 million) with 3.2 billion RMB spent on R&D and a net loss of 4.7 billion RMB [3]. Its share price surged from 116.20 HKD at IPO to around 1,100 HKD by August 2026 [3].
Usage metrics differ somewhat: OpenCode reported approximately 42 trillion tokens processed by Ox Alpha over six days, while OpenRouter cited 27.25 trillion tokens in one week [6, 8]. Z.ai guarantees the model does not use user prompt data for training, ensuring privacy [7].
The free trial period for GLM-5.3-Flash is expected to end around August 31. Z.ai has yet to announce pricing beyond the current token costs. The company plans to release detailed financial results for the first half of 2026 on August 31 [1, 3, 6, 7].
