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

TL;DR Highlight

Digging through Muse logs turned up one session likely using an Azure-served GPT model alongside Avocado.

Who Should Read

Backend developers designing agent runtimes that route across model providers, or verifying routing through logs.

Core Mechanics

Only one session used a different model
- Most session logs in the author's VM were routed to Meta's internal Avocado model.
- A single subagent session used azure/muse-special.
OpenAI-style traces showed up
- The signature was tagged gpt_responses_v1 with an encrypted payload starting with gAAAAA.
- Tool call IDs were call_ plus 24 mixed-case characters, versus call_ plus 32 hex characters in Avocado sessions.
The repo had clues too
- A repo search found the line "GPT Responses model client via MAGI native Azure OpenAI lane."
- The model catalogue lists azure/muse-special right before azure/gpt-5.6-sol.
The exact model is unknown
- Files and logs don't show which GPT model it is or why the subagent picked it.
- The author's best guess is an OpenAI model served through Azure.
The catalogue lists many providers

Community Reactions

Opinion that the title is misleading
- One commenter saw no evidence it's an OpenAI model and called the title very misleading.
Doubt about why they'd route elsewhere
- A commenter wondered why Meta would route to other models, since it can't be a compute shortage.
- They admitted it may not make sense and they could be missing something.
Curiosity about public filings
- One commenter asked whether a Meta-OpenAI deal would show up in official filings.
Experience with odd Muse 1.3 spark output
- A commenter said it occasionally returns Chinese text that looks like internal instructions.

How to Apply

Log the model ID on every agent call
- Today, record both the requested and returned model name in each session log.
- Aggregate model distribution per session to spot unexpected models.
Compare tool call ID formats by provider
- Compare tool call ID patterns (length, character set) across providers in your logs.
- Pull out sessions with a different format and check which route they took.
Check privacy terms before adopting an external agent
- Confirm in the terms whether conversations can be used for model development and how to opt out.
- Assume server-side routing can change and limit what sensitive data you send.

Try It

copy & run

# Paste into Claude Code or a terminal agent (simplified version of the article's method)

Analyze my agent session log directory {{log path}}.

1. Count the model ID used per session and show a table.
2. Pull out any sessions that used a model different from the majority.
3. Compare those sessions' tool call ID format (prefix, length, character set) with the others.
4. If there are signature or encrypted fields, tell me their prefix patterns.

Separate confirmed facts from guesses.

Next step: turn the format differences into a logging rule that aggregates automatically every session.

Terminology

AvocadoThe name of Meta's internal model. Most sessions in the author's VM were routed to it.

DistillationTraining one model on another model's outputs. Like studying from a senior's answer sheet.

RL (reinforcement learning)Improving a model by rewarding good outcomes. Like trial and error in a game to raise the score.

Responses APIOpenAI's response-generation API. In the article the gpt_responses_v1 signature was cited as a trace of it.

Kill switchA switch that instantly turns a feature off when something goes wrong. The article says an env setting comment calls it that.