Ox Alpha, a stealth model with a 1,048,576-token context window that appeared on public endpoints in August, still has no named creator, and a paper published the same week explains why nobody outside can settle the question. The model's own about page describes a reasoning model accepting text, image, and video with tool and JSON support, and names nobody. Z.ai's GLM 5 documentation lists it as a preview model without stating what it is. The identification methods people are reaching for do not work at this distance.
Key facts
- Ox Alpha: 1,048,576-token context, 131,072 maximum output tokens, text, image and video inputs, no disclosed creator.
- Its self-published benchmark page claims 8 of 10 on a small real-world coding set against GLM-5.3 at a 62 percent mean pass rate, and explicitly calls the sample small and directional.
- The lineage paper, arXiv 2608.14929 by Aman Singh Thakur and Rayan Khoury, reports perfect separation on its controlled benchmarks and runs 76 times faster than the nearest robust baseline on GPT-2.
- Primary sources: the Ox Alpha site and Business Insider's August 22 report.
The paper is called "Training Leaves Traces," and its method is genuinely elegant. Every transformer block adds a computed update on top of a running signal, the residual stream. The authors strip out the part of each block's contribution that just passes the input through unchanged and keep the remainder, which is the part specific to what that block actually learned. Compare those remainders across blocks, match them up between two checkpoints, and you get a symmetric score that separates related models from unrelated ones. In a public case study on LLaMA-2 checkpoints it cleanly separated three related models from seven unrelated ones.
The catch is in the access requirement. This is white-box: you need the weights of both models, and the architectures have to be compatible enough to line the residual blocks up. Nobody has Ox Alpha's weights. All anyone has is a chat endpoint. So the strongest available provenance method in the literature is exactly the wrong tool for the question the internet is asking, which is a useful thing to be clear about, because the paper is being cited in the argument as though it settles it. It relates to model fingerprinting, but fingerprinting from behavior alone is a different and much weaker discipline.
How weak became measurable this same week. The reasoning-trace paper found that prefilling one model with the first one percent of another model's reasoning trace shifted its visible answer style toward the source on 29 out of 30 problems. Read that carefully: a small prompt-level intervention makes model A start sounding like model B on nearly every problem. If style resemblance is that easy to induce deliberately, then observing style resemblance in the wild tells you very little about who trained what. The authors of that paper say the same thing about their own most eye-catching result, calling it suggestive rather than proof.
The public argument has not absorbed that. Business Insider's August 22 report described early speculation pointing at a Chinese lab, with a competing theory tying it to Microsoft, and the piece reported the dispute rather than resolving it. One independent analysis argues from tokenizer behavior, video-token consumption, and response-style probes that Ox Alpha is likely a next-generation multimodal model in the GLM family. That is a hypothesis built from tests, and it is a reasonable one, but the underlying social-media threads it draws on were not directly retrievable for verification, and inference from probes is not identification.
The reason this matters beyond curiosity is that anonymous frontier-capable models are becoming a category rather than a stunt. Ground Truth covered a free million-token model that appeared with no owner and two privacy policies and DeepSeek selling a checkpoint it has not published. If a model can serve a million tokens of context to anyone with an API key while its provenance, training data, safety testing, and legal jurisdiction are all unknown, then every downstream question about accountability has no addressee. That is a governance problem wearing a mystery-model costume.
The honest caveat runs in both directions. Ox Alpha's capability claims are self-published: its own benchmark page reports 8 of 10 on a ten-task coding set against GLM-5.3's 62 percent, and to the site's credit it labels that sample small and directional. Nobody should treat it as an independent result. And the only ownership statement anyone can currently stand behind from primary sources is the boring one: the model is publicly reachable, and it is officially anonymous. Everything past that is inference. Given how easy this week's research showed it is to make one model wear another's voice, inference should carry less weight than it currently does.
Originally published on Ground Truth, where every claim is checked against the primary source.
