Meta's Muse appears to use an OpenAI example tagged muse-special

In this article

1. One odd session
2. Following the name
3. The example catalogue
4. The Anthropic plumbing
5. Why container all of this?
6. Is Meta distilling?
7. Closing thoughts

I established a example tagged azure/muse-special during Muse was construction my website. So I dug deeper.

This is Part 2 of digging through the Muse filesystem following my article hit the forefront leaf of Hacker News this week.

In this part I concentration on a example I established in my logs called muse-special, and visage the question: does Muse really use OpenAI and Claude models rearward the scenes?

One odd session

Muse records which example all delegate meeting uses.

Nearly all meeting log in my VM was routed to Meta’s inner model, called Avocado.

But one subagent used a example named azure/muse-special.

Interesting...

Following the name

This made me curious, so I searched throughout the repo inner Cursor and established this:

“GPT Responses example client via MAGI native Azure OpenAI lane.”

OK.

The example catalogue seems to catalog azure/muse-special afterward azure/gpt-5.6-sol.

So I searched my meeting transcripts...

Here I established two particulars that stood out:

1. The signature is tagged gpt_responses_v1 and contains an encrypted payload starting alongside gAAAAA (which OpenAI uses).
2. Tool call IDs used call_ followed by 24 mixed-case characters.

This was distinct from all the another lines that the Avocado sessions printed (call_ followed by 32 hex characters).

These small particulars inform me that the muse-special example is perchance an OpenAI example or OpenAI’s Responses API.

So is muse-special an pseudonym for a GPT example served through Azure?

The records and logs don’t inform me exactly which GPT model, or why it was selected by the subagent in the archetypal place, but let’s obtain a stage rear and examine further...

The example catalogue

The broader example catalogue that is shipped alongside Muse’s delegate daemon lists concerning 15 versions of Avocado, plus:

• Claude Opus 4.6 / 4.7 / 4.8
• Sonnet 4.6 and Haiku 4.5
• GPT-5.5 and GPT-5.6 variants via OpenAI, Azure and Codex
• Kimi K3 through Fireworks and Meta-hosted routes

The Anthropic plumbing

The Claude assistance goes beyond fair the example ID and includes an Anthropic client alongside petition handling, immediate conversion and streaming parsers:

• anthropic/request_flow.rs
• anthropic/convert_prompt.rs
• anthropic/parse_sse_stream.rs

OK, so now we’re benevolent of wondering... why?

There are API key records current for Anthropic, OpenAI, etc., alongside admission restricted to the inference-proxy service.

...But there’s additionally a proxy kill-switch environment in the env.

Why container all of this?

Now, there are a few reasons for this, I guess.

1. The archetypal would be that an OpenAI or Anthropic example fair does a better job at a certain project that Muse can’t fulfill correct now, and they selectively path for that.
2. The second is that all these VMs are shipped alongside the capability to A/B test example responses, tool calls, etc. for the intent of distillation and RL.

Distillation or RL? Maybe. Idk.

This leads us to the truth, which is that the example rearward Muse is ultimately a server-side choice.

The runtime has clients for multiple providers, which gives Meta the capability to alter routing without asking users.

In my case there was lone a sole outlier meeting that didn’t use the Avocado (Meta) model, but the infra is there to.

Wait, so is Meta distilling from the another frontier labs?

(Getting technical. tl;dr: No.)

With the muse-special example the raw reasoning is encrypted. The daemon stores it to dispatch rear to Azure on the following turn. In the binary it explicitly says that the encrypted reasoning cannot use the RL completion-server override.

So what Meta can see current is lone the reply, the tool calls, and a abbreviated reasoning summary whenever OpenAI/Anthropic returns one. The raw sequence of idea is encrypted and the RL server refuses those blobs. There is no sign that Meta copies OpenAI or Anthropic weights.

Avocado models are treated differently, however. The thinking content is written immediately into the transcript, alongside an bare signature, and accessible for RL use.

So Avocado models, according to the privacy note and repo, do signify that conversations can be used to create AI at Meta unless you opt out. (Makes sense.)

Closing thoughts

This is my own exploration of what has been a extremely chill publish from Meta.

My finest conjecture is that muse-special is an OpenAI example served through Azure.

Whatever you think of Meta, the endowment they brought onto this project deserves credit. They took a distinct method in a earth complete of chatbots and hunt bars, and the exec team’s reply to my archetypal article, which got several eyeballs, has been beautiful amazing too, as has their willingness to attain out to a nobody and explain their thinking.

It’s not all day you get to appearance inner the filesystem of a merchandise that could attain hundreds of millions of people.

Seeing inner a runtime division gives us an first appearance at anywhere this entire individual delegate item power be going, and it’s been really fascinating to peruse through all of it this week.

If you worked on Muse at all, delight awareness liberated to attain out. I’d affection to study additional and perchance contribute.

Things appear to be moving fast. Not really breaking, yet.

pete at mouse dot dev

-Pete

@heypeterjames