Forensic Analysis Suggests Anonymous Model ox-alpha Is Zhipu's GLM-5.3-V

On August 22, researcher zainhas published a series of forensic analyses arguing that the anonymous test model ox-alpha is in fact Zhipu's GLM-5.3-V. Combining four lines of evidence—tokenizer, special tags, metadata, and prompt forensics—the conclusion gained significant traction in the community, though some cautioned that parts of the evidence are of limited strength.

Confirmed

• Tokenizer exact match: using a preregistered blind-test method, ox-alpha's tokenization counts for 9 unique special strings exactly matched GLM, while other models diverged massively
• Special tag identification: when asked to complete text containing the GLM-specific special tag <sop>, ox-alpha refused and said completion would reveal its identity, then identified itself as a GLM-series large language model, while denying it was developed by Zhipu AI, Moonshot, or DeepSeek
• Chinese prompt leak: forcing the model to reason in Chinese with Chinese prompts caused information leakage in 7 out of 12 samples, with replies identifying itself as a 「General Language Model」(GLM) series; one sample directly gave the registered entity name 「北京智谱华章科技」
• Metadata match: ox-alpha's metadata aligns with GLM5.x, including inference hierarchy, context length, and sampling parameters

Unconfirmed

• @Afinetheorem noted that although using Pliny's system prompt can make ox-alpha claim to be GLM, this may simply be a token pattern left over from training data rather than information embedded by the developers in a system prompt, so self-report-type evidence alone cannot prove the model's underlying architecture is GLM
• Earlier, based on cl100kbase tokenizer traits, the community had speculated ox-alpha was Microsoft's unreleased MAI-2 frontier model (a guess that excluded OpenAI, Google, Anthropic, xAI, and Chinese frontier labs), which contradicts the latest tokenizer comparison conclusion—whose is right awaits an official response

Why it matters

• Tracing the identity of stealth test models bears on the fairness of community benchmarks: if anonymous models can be identified through tokenizer and prompt forensics, vendors' stealth-release strategies become effectively pointless
• This analysis demonstrated a multi-evidence cross-validation forensic methodology (preregistered blind tests, special tag probing, metadata comparison), providing a reusable template for identifying anonymous models in the future

2026-08-22 ~ 2026-08-22 · 12 related posts

• Ox Alpha Speculated to be Microsoft's Unreleased MAI-2 Frontier Model — inductionheads · 2026-08-22
• Prompting ox-alpha to identify as GLM proves little about the model — Afinetheorem · 2026-08-22
• Test shows ox-alpha denies being developed by Zhipu, Moonshot, or DeepSeek — zainhas · 2026-08-22
• Evidence suggests ox-alpha is GLM-5.3-V — zainhas · 2026-08-22
• ox-alpha tokenizer matches GLM exactly in blind test — zainhas · 2026-08-22
• [source] Technical confirmation: Tokenizer analysis proves ox-alpha is GLM-5.3-V — zainhas · 2026-08-22
• [source] Forensic deep dive: Chinese reasoning forces ox-alpha to leak Zhipu entity name — zainhas · 2026-08-22
• Model identity leak: Agent reveals itself as GLM via special tokens — zainhas · 2026-08-22
• ox-alpha metadata matches GLM5.x exactly — zainhas · 2026-08-22
• [source] Reverse engineering confirms mystery model 'Ox Alpha' is Zhipu's GLM — bdsqlsz · 2026-08-22
• Ox Alpha Mystery Deepens: Speculation Points to a Post-Trained Chinese Model — brandon_galang · 2026-08-22
• Rumour: mystery model Ox Alpha on Hermes is Zhipu's GLM series — AJChadha · 2026-08-22