Z.ai’s Ox Alpha model, the anonymously tested version of GLM-5.3-Flash, processed approximately 23 trillion tokens over six days on OpenRouter before its Beijing-based creator confirmed its identity last week, making it the biggest model launch in OpenRouter’s history at roughly 2.3 times the volume of the next closest model.
The company confirmed on 26 August 2026 that Ox Alpha was GLM-5.3-Flash, its latest open-weight release, ending days of speculation among developers who had been guessing whether a major Western lab or a Chinese rival was responsible for the high-performing anonymous model.
How the Z.ai Ox Alpha model seized the spotlight
Z.ai began the anonymous test on 20 August 2026, listing the model on OpenRouter and OpenCode under the Ox Alpha name. On OpenCode, the model ended DeepSeek’s 56-day run at the top of the leaderboard. Stripe chief executive Patrick Collison called it ‘very impressive’, a comment that landed with additional weight given that Stripe had agreed to acquire OpenRouter the day before the model appeared.
The mystery spread to Chinese social media too. Users nicknamed it the ‘Niu Lai Model’, a reference to a low-budget animated film about a calf that had become a meme phenomenon that month.
Z.ai confirmed the model’s identity to Bloomberg on 26 August. Five hours later, OpenRouter’s production catalogue listed z-ai/glm-5.3-flash, and Z.ai published its MIT-licensed weights on Hugging Face the same day. Its shares closed more than 12% higher the following session, trading at roughly ten times January’s IPO price.
This was not Z.ai’s first anonymous experiment. The company had previously tested GLM-5 under the name Pony Alpha using the same approach of gathering user feedback before a public release.
What GLM-5.3-Flash can do
According to the official model card on Hugging Face, GLM-5.3-Flash carries 320 billion total parameters with 18 billion active, was trained on a 30-trillion-token multimodal corpus, and offers a context window of 1,048,576 tokens. It uses a hybrid sparse and linear attention architecture that reduces KV cache size fourfold compared with GLM-5.3, with three times the serving performance. On OpenRouter it is priced at $0.075 per million input tokens and $0.25 per million output tokens.
The model is multimodal, handling text, images, and video. Z.ai said all traffic from the Ox Alpha test ran on Chinese AI chips, compensating for limitations in memory capacity and bandwidth through software optimisation and workload distribution across hardware, a constraint shaped by Washington’s restrictions on advanced chip exports to China.
The company behind the model: Z.ai’s finances and origins
Z.ai, known in China as Zhipu AI, was founded in 2019 using technology developed at Tsinghua University in Beijing. It listed in Hong Kong on 8 January 2026 under the corporate name Knowledge Atlas Technology, later rebranding its English name to Z.ai in July. CGTN reported it became the world’s first listed large language model company upon that listing.
The IPO priced shares at HKD 116.20 each. Baker McKenzie advised on the offering, which raised over HKD 4.3 billion (approximately USD 560 million) across 37,419,500 shares at an implied valuation of roughly HKD 51 billion (USD 6.55 billion). Shares now trade at around HK$1,100, approximately nine times the IPO price.
The company’s 2025 full-year revenue of RMB 724.3 million more than doubled from the prior year but fell short of the mean analyst estimate of RMB 760 million polled by Reuters, according to CNBC. The net adjusted loss widened to RMB 3.18 billion, up 29.1% year-on-year, as R&D spending climbed.
The annual results filed with HKEx for the year ended 31 December 2025 show R&D expenses of RMB 3.180 billion, up 44.9% from RMB 2.195 billion in 2024, and gross profit of RMB 296.656 million, a 68.7% increase on the prior year’s RMB 175.889 million.
By revenue mix, on-premises deployment accounted for 74% of total revenue in 2025, according to Yicai Global. The enterprise AI agent segment grew 249% and the open platform and API business nearly quadrupled to RMB 190 million.
Z.ai competes against DeepSeek, Moonshot AI and Alibaba’s Qwen team on performance and price under tight computing constraints. The company was added to the US Commerce Department’s Entity List in January 2025 over alleged links to China’s military, a designation it has continued to operate under as it pursues international expansion.
Z.ai is scheduled to report results for the first half of the year shortly. Whether GLM-5.3-Flash’s record-breaking OpenRouter debut translates into a step-change in API revenue will be the number investors watch most closely.