Ox Alpha · Community source · Editorial analysis
The article compares Ox Alpha with public GLM, Gemini, DeepSeek, Kimi, and MiMo using tokenizer behavior, video-token budgets, and server errors, strongly pointing to a GLM-family model on a Z.ai serving stack while explicitly stopping short of naming a product or developer.
Unverified: the original source could not be rechecked. Historical figures below are not current verified results.
The article compares Ox Alpha with public GLM, Gemini, DeepSeek, Kimi, and MiMo using tokenizer behavior, video-token budgets, and server errors, strongly pointing to a GLM-family model on a Z.ai serving stack while explicitly stopping short of naming a product or developer.
Object: OpenRouter stealth/ox-alpha.
Input specification: 1,048,576-token context, text/image/video input, and text output; the public model card lists an anonymous Stealth provider.
Data sources: Measurements attributed to YFarmX, unclecode, Ben Davis, and Security Kid; the author explicitly says they were not rerun in this article.
Comparisons: Public GLM-5.3, GLM-5V-Turbo, Gemini, DeepSeek, Kimi, and MiMo.
| Measurement | Ox Alpha comparison | Interpretation boundary |
|---|---|---|
| Prompt tokens on 50 strings | 50/50 matches GLM-5.3; Gemini 11/50, DeepSeek 9/50, Kimi 9/50, MiMo 18/50 | Inherited community measurement; tokenizer match is not weight or product match |
| Video-token budget | Four clips match GLM-5V-Turbo at 296, 296, 884, and 1,064 respectively | n=4; encoder/budget similarity does not prove the same model |
| Server errors | The article says Java/Spring errors and business codes 1210/1214 are byte-identical to public GLM-5.3 errors | Shared serving stack is a hosting clue, not proof of weight ownership |
| Product-card comparison | Ox is 1M multimodal; public GLM-5.3 is 1M text-only, while GLM-5V-Turbo is 200K vision | “Same family” and “same public SKU” must remain separate |
The article's value is the reproducible numbers and explicit boundaries, not the claim that Ox Alpha is a named model. The stronger current conclusion is a GLM-style tokenizer/video encoder and a Z.ai PaaS hosting clue; the exact checkpoint, operator, and developer remain undisclosed.
Fix the 50 strings and tokenization method, query Ox Alpha and public models, and save prompt_tokens.
Fix the four video clips, dimensions, and frame rates; record visual-token or server-reported budgets and expand to far more than four samples.
Save raw HTTP errors, response headers, timestamps, and model versions; compare error bytes rather than screenshot similarity.
Score tokenizer, video budget, server stack, and product card separately. Update the identity conclusion only after a first-party claim, weight match, or larger-sample evidence appears.
The article says the measurements came from other people's work and were not rerun by its author.
Public model cards, anonymous-provider data, and community fingerprints involve inference; “GLM family” must not be rewritten as “GLM-5.3 Flash confirmed.”
The article's state is as of 2026-08-25; the server, model card, and anonymous preview can change.
The figures, task set, reasoning tier, and client conditions apply only to the listed source and collection snapshot. Different versions, harnesses, or providers must not be compared directly; undisclosed parameters remain unknown.
For a reproduction, fix the model version, provider or client, reasoning tier, tools, task-set version, sample count, and collection date, and record failures, retries, and human corrections. Full steps are in the source notes below.
X Article · Jonathan Turner (@JonathanNTurner), with David and the house · Original publication date 2026-08-25 · Site edit date 2026-09-20
Open original sourceOx Alpha
Download the Tabbit client to check model access
