modelprint fingerprints any OpenAI-compatible endpoint in the browser, and pinned OpenRouter's stealth Ox Alpha to the GLM family

GitHub·medium signalXBlueskyLinkedInCopy link

unclecode/modelprint runs 9 infrastructure probes against any OpenAI-compatible endpoint from a static page with no server, keys never leaving the tab. Its day-one run against 12 candidates scored the stealth model stealth/ox-alpha at 6 of 9 probes and 4 of 4 normalized tokenizer counts matching z-ai/glm-5.3, while every other lab's best was 2 of 4. The author built it after showing that DeepSeek's API silently serves a different model under the old deepseek-chat alias, which is the general case this checks for.