Meta's Muse appears to use an OpenAI model labeled muse-special

Meta's AI website builder may use OpenAI models, blurring the lines of proprietary AI development.

An investigation into Meta's recently released Muse AI website builder has revealed that one of its subagents, identified as 'muse-special,' may be utilizing an OpenAI model. The author discovered logs showing this subagent connecting to 'azure/muse-special' instead of the expected Meta-developed 'Avocado' model. Further analysis of session transcripts and file system data uncovered signatures like 'gpt_responses_v1' and encrypted payloads starting with 'gAAAAA,' both indicative of OpenAI's API. The model catalogue shipped with Muse also lists variants of GPT models through OpenAI and Azure, alongside Anthropic's Claude and Meta's own Avocado. While Meta's infrastructure includes clients for multiple AI providers, the logs suggest 'muse-special' is distinct from Meta's internal models, leading to speculation that Meta is routing specific tasks to OpenAI's models via Azure for potentially superior performance or for future model distillation and reinforcement learning.

The presence of 'muse-special' routing to an Azure OpenAI endpoint, coupled with signature details matching OpenAI's GPT responses, strongly suggests Meta is leveraging external LLMs for specific functionalities within Muse. This move allows Meta to offer advanced capabilities without necessarily developing them in-house, potentially accelerating product development and offering users access to best-in-class models for certain tasks. The infrastructure's flexibility to route to various providers also hints at a strategy of A/B testing and future model distillation.

This discovery has implications for the competitive landscape of AI-powered development tools, signaling a trend where companies might integrate third-party frontier models to enhance their offerings rather than solely relying on internal development. It raises questions about Meta's long-term AI strategy and its commitment to open-sourcing or proprietary development, while also benefiting users with potentially more powerful and versatile tools. Future observation should focus on whether Meta expands its use of external models or shifts towards greater internal development and distillation.