Meta Connect: Personal Superintelligence on a Keychain

# The Hardware

Muse Charm is the first consumer product built on Meta's third-generation Training and Inference Accelerator (MTIA-3). The chip was designed from the ground up for transformer inference at edge power budgets. At 3 watts, it achieves roughly 15 tokens/second on the 3.6B active parameter model — comparable to what an M2 MacBook Air does with a 7B model, but in a package smaller than an AirTag.

# How the 72-Hour Context Works

Muse Charm continuously records ambient audio, runs real-time speech-to-text on-device using a distilled Whisper-v4 model, and builds a rolling 72-hour context window. Audio is never stored raw — only the transcribed text and extracted semantic embeddings persist. After 72 hours, old context is compressed into summary nodes using a separate summarization pass.

The result is that you can ask "What did Sarah say about the project timeline in yesterday's meeting?" and the device will retrieve the relevant transcript segment, generate an answer from context, and respond through its bone conduction speaker — all without touching a network.

Audio In (3-mic array, noise-cancelled)
  → Whisper-v4-tiny (real-time STT, 0.8W)
  → Speaker diarization (who said what)
  → Semantic embedding (128-dim per utterance)
  → Rolling context store (72h window, ~2GB)

Query ("What did Sarah say about timelines?")
  → Embedding search over 72h context
  → Top-K transcript retrieval
  → Llama-5 Scout inference (3.6B active, ~15 tok/s)
  → Answer via bone conduction speaker

# How It Compares

Muse Charm is not the first AI wearable. But it is the first with genuine on-device language model inference. Here is how it stacks up against the current field.

The key differentiator is clear: Muse Charm is the only wearable that can reason about your day without calling home. Every other device is essentially a microphone that sends audio to a cloud API. Muse Charm runs the model on your keychain.

# The Privacy Questions Nobody Asked

At the Connect keynote, Mark Zuckerberg spent 14 minutes on Muse Charm and zero minutes on privacy. The press Q&A was limited to five questions, all about hardware specs. Here are the questions that should have been asked.

# What Developers Should Know

Meta announced a developer SDK launching in January 2027 alongside the hardware. Muse Charm will support third-party "Skills" — lightweight app-like experiences that run on-device using the Llama-5 Scout model and local context.

name: "Meeting Summarizer"
trigger: "summarize today's meetings"
context_access:
  audio_transcript: true
  calendar_events: true
  contacts: false
cloud_escalation: optional
output: text  # spoken via bone conduction
max_tokens: 500
permissions:
  - ambient_context_read
  - calendar_read

Skills run sandboxed on the MTIA-3 chip. They can read from the context store but cannot write to it or access raw audio. Cloud escalation is opt-in per skill and requires user approval at install time. Meta says the Muse SDK will be open-source, with skill distribution through a curated store (not sideloading — at least initially).