mirror of
https://github.com/hanzoai/ai.git
synced 2026-11-20 14:13:23 +00:00
Every claim is a real request against inference.do-ai.run, and every repro in it was re-run against the live API before filing. The raw per-model audit (68 models, what each actually returns) is alongside it as do-model-audit.json so the findings can be re-checked without re-probing. Four findings: context_length wrong in both directions (deepseek-v4-pro serves 1M against an advertised 87K; nemotron serves 262K against an advertised 1M); 35 of 68 models 403 'not available for your account' while the console lists them as available Serverless; fal-ai/* and ideogram-* exist in one surface but not the other; and qwen3-tts-voicedesign fails inside DO's gateway on requests that pass DO's own schema validator.