An investigation into the filesystem of Meta's Muse agent has uncovered evidence suggesting the platform maintains the capability to route tasks to third-party AI models. While the vast majority of agent sessions are handled by Meta's internal model, referred to as 'Avocado,' researchers identified a specific session log utilizing a model labeled 'azure/muse-special.' Technical indicators, such as OpenAI-style tool call identifiers and encrypted payloads consistent with OpenAI's response formats, suggest that this specific model may be an OpenAI variant hosted via Microsoft Azure. The Muse agent runtime also includes a comprehensive model catalogue that lists various versions of Claude, GPT, and Kimi models, alongside infrastructure for Anthropic request handling. The presence of these clients and associated API keys indicates that the system is architected to support multi-provider routing. This architecture allows Meta to dynamically select different models for specific tasks, potentially to leverage superior performance in certain domains or to facilitate A/B testing for model distillation and reinforcement learning. However, the exact criteria for this routing remain opaque, and there is no evidence that Meta is copying weights from these external providers. The raw reasoning chains from these third-party models are encrypted and inaccessible to Meta's internal reinforcement learning servers, distinguishing them from the Avocado model, which integrates directly into Meta's development pipelines.
The article discusses how researchers identified a specific session log labeled azure/muse-special, which provides evidence of third-party model routing within the Muse agent.
The Muse agent runtime includes a comprehensive model catalogue that lists various versions of Claude, GPT, and Kimi models, indicating support for multi-provider routing.
The presence of specific clients and associated API keys, such as the Anthropic reverse proxy override, confirms that the system is architected to support multi-provider routing.