Free and powerful! The mysterious model Niu Lai has sparked heated discussions online, with speculation about whether it is Z.AI or someone else entirely.

A model without a signature is sending the entire AI developer community into a guessing game. An anonymous AI model named Ox Alpha has quietly appeared on the OpenRouter platform, available for free and boasting strong performance, quickly sparking discussions among developers. Domestic netizens colloquially dubbed it The Cow Has Arrived model, as Ox means "cow". Free and powerful, developers flocked to it. Late night on August 20, Eastern Time (early morning August 21, Beijing Time), Ox Alpha was quietly launched on OpenRouter as a stealth model by an anonymous third party. The specifications are quite luxurious: a context window of 1 million tokens, support for multimodal input including text, images, and videos, full support for tool calls, and forced inference activation. The price is even more outrageous completely free. According to Bloomberg, Ox Alpha was open for free for a week, with usage described as almost unlimited. The AI programming tool OpenCode announced that its provider possesses daily computational power of 100 trillion (100T) tokens about 100 times the monthly AI token usage of Visa. Patrick Collison, CEO of Stripe, personally tested it and tweeted: Very impressive. Coding tests: surpassing GPT-5.6 and Claude. Davis7, author of the AI programming framework fable, ran 10 questions through DeepSWEs official evaluation, and the results left him stunned. He wrote on X: I am very confused. The data speaks for itself: Ox Alpha scored 80%, Claude Fable 65%, and GPT-5.6 Sol 52%. An anonymous model has outperformed two well-known closed-source leading models by 15 percentage points. Subsequently, someone in the community conducted independent re-tests, and the conclusions were largely consistent: 4 passed, 4 close to passing, 2 failed, matching the 8/10 result. Two independent tests pointed to the same conclusion: this mysterious models coding capability is indeed at the cutting edge. A 24-hour fingerprint detective battle: five pieces of evidence point to Z.AI. With the identity a mystery, the community began to dig it up. According to a detailed analysis by Lizikk Zhu, founder of Atomstorm.ai and former executive at Baidu/Alibaba, the community was able to identify the main suspect within 24 hours through technical fingerprint comparison of public APIs. The strongest evidence comes from the video encoder. Feeding the same set of controlled test videos into different models and recording their respective token consumption revealed that Ox Alpha's usage matched precisely with Z.AIGLM-5V-Turbo: 296, 296, 884, 1064, with no discrepancies. This analysis noted that it aligned with three independent design features: frame sampling strategy, duration scaling, and resolution scaling, with a probability of coincidence being virtually zero. Four additional pieces of evidence followed: Text tokenizer: 25 special strings (including CJK, emojis, Cyrillic letters), with token counts identical to GLM-5.3. Error code: returns [1210] error when incorrect parameters are passed this code is unique to Z.ai according to the analysis. Chinese backend: returns an error in Chinese when image parsing fails: Image input format/parsing error. Censorship responses: provides official responses to sensitive questions. These five pieces of evidence come from completely different technical aspects but all point to Z.AI. Lizikk Zhu concluded that Ox Alpha is likely the unreleased GLM-5.3 vision version (or GLM-5.5) from Z.AI, with a confidence level of 70-80%. According to Business Insider, tech media such as Wccftech also highlighted that Z.AI had previously conducted anonymous testing of GLM-5 under the codename Pony Alpha, and developers found similar tokenizer behaviors and response patterns in Ox Alpha. However, the mystery may not necessarily be Z.AI. Doubts also exist. First, the computational power doesnt add up. With a free operation of 100T tokens per day for a week, multiple analyses suggest that Z.AIs current computational reserve is unlikely to support this independently. The community thus proposed the synthetic hypothesis: GLM weights + major power provider hosting the model belongs to Z.AI, but the server might be Xiaomi or Tencent. Second, someone has hinted at a reversal. AI insider Mia posted on X: I have confirmed whose Ox Alpha is, but I cannot say now. You will be very excited its not what you think. This post immediately garnered a multitude of question marks. According to Business Insider, AI analyst Andrew Curran noted on X that GLM was still the mainstream speculation on Friday evening, but by Saturday morning, everyone seemed less certain about anything. Wccftech later added another analysis: the tokenizer features of Ox Alpha might also point to Microsofts MAI series models. The mystery is likely to be revealed on August 28 the day when Z.AIGLM-5.3 weights are officially open-sourced, and once the model card is released, its identity may become clear. Where did the data go? Retention policies contradict each other. Aside from speculating identities, a more practical question remains unresolved: the fate of user data. According to Lizikk Zhu's analysis, the OpenRouter model card clearly states, Prompts and completions are retained by the provider, not used for training. Yet, OpenCode's official announcement claims, Zero Data Retention. Two official channels, two statements, at least one of which is incomplete. Not used for training and zero data retention are two different concepts: the former implies that conversations are still in the provider's hands, with an opaque retention period; the latter only means that requests end at deletion and can be audited from a compliance perspective. Anonymous launches have become industry practice. According to Bloomberg, this is not an isolated phenomenon. Companies like ByteDance, Sea, and Alibaba have all released models this year without disclosing their sources, allowing for testing new services with fewer constraints. Lizikk Zhus timeline shows that this script has several precedents: Hunter Alpha was later confirmed to be Xiaomi's MiMo-V2-Pro, Owl Alpha was Meituan's LongCat 2.0, Pony Alpha was Z.AIGLM-5, and even OpenAI once tested GPT-4.1 under the codename Quasar Alpha. The logic behind this is clear: anonymous launches circumvent brand risks, trading free access for real production traffic, and stress-testing feedback is more valuable than internal testing data, ultimately leading to monetization upon claiming a public release. If Ox Alpha ultimately earns an AA intelligence index score of 61 as predicted by analysts, it will become the first open-source weight model to break the 60-point barrier, directly touching the threshold of closed-source cutting-edge technology. This article is reproduced from the Wall Street Watch app, author: Long Yue; GMTEight editor: Song Zhiying.