ModelPrint — Fingerprint Any OpenAI-Compatible Endpoint

• URL: https://github.com/unclecode/modelprint
• Date Saved: 2026-08-24
• Source: GitHub
• Tags: ai-engineering, ai-tools
• Repo: https://github.com/unclecode/modelprint (77 ⭐, created 2026-08-21)

Summary

Browser-based tool to fingerprint which model/lab is actually behind any OpenAI-compatible API endpoint. No install needed — runs entirely in-browser, keys never leave the tab.

Why it exists

• Built after “Ox Alpha” stealth model appeared on OpenRouter (Aug 21 2026) and community turned detective
• Also caught DeepSeek silently swapping models behind deepseek-chat alias
• Core thesis: “Plumbing does not lie; personality does” — tokenizers and error handlers reveal identity, personality/censorship tests don’t

9 Infrastructure Probes

1. Tokenizer counts (English/Chinese/Code/Emoji) — normalized against 1-char baseline so host templates cancel out
2. Template offset — prompt-token overhead reveals serving template size
3. Temperature 2.0 error — validation messages are written by lab engineers
4. max_tokens overflow — refusal names the real output limit
5. Error code family — GLM’s 1301 code gave away Ox Alpha
6. Finish vocabulary — finish_reason values differ per lab

Community probes (network forensics)

• Router region/provider metadata
• Generation ledger (native token counts)
• Header DNA (cf-ray vs x-amzn-requestid etc.)
• Context ceiling bisection
• Logprob geometry (quantization detection)

Key design decisions

• No censorship/personality probes — community proved these are unreliable (opposite conclusions from same data)
• Runs twice per tokenizer probe, reports “unstable” if router spreads across hosts
• Side-by-side comparison UI
• Community-extensible: add a probe by writing one JS file

Day 1 result on Ox Alpha

GLM family (z-ai/glm-5.3) matched 6/9 probes + 4/4 tokenizer. Next best was 2/9. Case closed.

Built by @unclecode (also author of Crawl4AI, 29K+ ⭐).