Meta's Muse Appears to Route to OpenAI via Azure — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Muse's session logs showed an outlier subagent using a model labeled azure/muse-special, whose signature (gpt_responses_v1, gAAAAA-encrypted payload) and tool-call format (call_ + 24 mixed-case chars) match OpenAI's Responses API rather than Meta's internal Avocado model.
- The model catalog shipped with Muse lists ~15 Avocado versions, Claude Opus 4.6/4.7/4.8, Sonnet 4.6, Haiku 4.5, GPT-5.5 and GPT-5.6 variants, and Kimi K3 routes, with API keys for Anthropic and OpenAI present and restricted to the inference-proxy service.
- The Muse daemon includes dedicated Anthropic client code (request_flow.rs, convert_prompt.rs, parse_sse_stream.rs), giving Meta a multi-provider runtime that can swap routing without user visibility.
- Reasoning from non-Meta models is stored only as an encrypted blob to relay back to the provider, and the RL completion-server explicitly refuses those blobs, so OpenAI and Anthropic conversations cannot feed Meta's training pipeline.
- Avocado (Meta's internal model) sessions store full thinking text directly into the transcript with an empty signature, making that reasoning usable for RL — unless the user opts out.
- The author, Pete (@heypeterjames), concludes that muse-special is most likely an OpenAI model served via Azure, though the logs don't specify which GPT variant or why that one subagent chose a non-Avocado route.
Why it matters: Meta shipped a consumer AI agent whose runtime quietly routes across OpenAI, Anthropic, and its own Avocado models without telling the user, and because non-Meta reasoning is encrypted, those conversations cannot feed Meta's training pipeline. That fallback lets Meta lean on GPT or Claude when Avocado can't handle a task, while only Avocado's plaintext traffic can improve Meta's own models.
Ask SkimNews
