Meta's Muse appears to use an OpenAI model labeled muse-special
3 hours ago
- Found a model labeled 'azure/muse-special' in logs, suspected to be an OpenAI model due to signature 'gpt_responses_v1' and encrypted payload starting with 'gAAAAA'.
- Tool call IDs for muse-special used 'call_' with 24 mixed-case characters, differing from Avocado's 32 hex characters.
- The model catalogue includes Avocado, Claude variants (Opus, Sonnet, Haiku), GPT-5.5/5.6 via OpenAI/Azure/Codex, and Kimi K3.
- Anthropic client files for request handling, prompt conversion, and streaming parsers are present, along with API keys for Anthropic and OpenAI.
- Possible reasons for using external models: selective routing for superior performance, or A/B testing for distillation/RL.
- Muse-special's raw reasoning is encrypted and cannot be used for RL, while Avocado's thinking text is available for RL training.
- Meta can change routing server-side without user notification, but there is no evidence of copying weights from OpenAI/Anthropic.