A fingerprinting tool pins Ox Alpha's backend to GLM

modelprint scored the stealth model listed as Ox Alpha at 6 of 9 infrastructure probes, the highest of 12 candidates checked on its first public run. Its tokenizer count matched z-ai's GLM-5.3 on all four probes, a result no other lab came close to. Every other model in the batch topped out at 2 of 4.

That gap matters for anyone paying for Ox Alpha access on the assumption it's a distinct model. A near-perfect match across infrastructure and tokenizer probes points to the same weights running behind a different name, not just similar architecture choices.

The tool checks any OpenAI-compatible endpoint from a static page, with no server and API keys that never leave the browser tab, according to modelprint's GitHub repo. Its builder wrote modelprint after finding that DeepSeek's API serves a different model under the old deepseek-chat alias, the same kind of mismatch the tool is built to catch.

Ox Alpha's operator hasn't said whether that tokenizer match means anything. If a repeat clean match shows up against the next batch of stealth entries, this stops being one flagged model and becomes a pattern.