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

In the wake of my article gaining front-page coverage on Hacker News, I delved further into the Muse filesystem. This investigation led me to a model labeled "azure/muse-special" in my logs. This prompted the question: does Meta's Muse actually utilize OpenAI and Claude models behind the scenes?

Upon examining the logs, I discovered that nearly every session was routed to Meta's internal model, called Avocado. However, one subagent utilized the "azure/muse-special" model, which sparked my curiosity. After searching across the repository, I came across the following line: "GPT Responses model client via MAGI native Azure OpenAI lane." This led me to believe that the "museum-special" model might be an OpenAI model or OpenAI's Responses API.

The model catalogue within Muse's agent daemon includes around 15 versions of Avocado, along with information about Claude support, Anthropic clients, and API key files for various providers, including OpenAI, Anthropic, and others. Additionally, there is a proxy kill-switch setting in the environment variables.

The runtime of Muse has clients for multiple providers, granting Meta the flexibility to modify routing without requiring user intervention. In my case, there was only one instance where Muse did not use the Avocado (Meta) model, but the infrastructure is present for this.

It appears that Meta does not copy OpenAI or Anthropic weights. Instead, the raw reasoning is encrypted and sent back to Azure on the following turn. Although the encrypted reasoning cannot be processed by the RL completion-server override, Meta can only see the reply, tool calls, and a short reasoning summary when OpenAI or Anthropic returns one. The raw chain of thought remains encrypted, and the RL server does not accept such blobs.

Avocado models, on the other hand, store the thinking text directly in the transcript, accompanied by an empty signature, making it available for RL use. According to the privacy note and repository, conversations involving Avocado models can be utilized for AI development at Meta unless opted out.

In conclusion, my analysis suggests that "museum-special" is an OpenAI model served through Azure. While Meta's approach raises questions, their talented team behind this project deserves recognition. They have taken a distinct stance in an industry dominated by chatbots and search bars. The executive team's response to my initial article was impressive, and their eagerness to engage with someone like myself, a mere observer, has been commendable.

It is rare to gain insights into the inner workings of a product that could potentially reach hundreds of millions of people. Exploring Muse's runtime cell has provided an early glimpse into the future of personal agents, and it has been an enlightening experience.