Muse MuseConcept
Muse MuseConcept

Independent concept built on Meta's Model API. Not an official Meta product.

Muse Privacy

Muse Muse builds the product. Muse Privacy is how it touches your private files: your device reads them, and Muse gets tags.

A connector for Muse, not a hosted service: it runs on hardware you own, and nothing on this site can reach your files. It hasn't been tried with real users yet, and this page says what has and hasn't been measured.

See one run

Recorded, not live. Scene rehearsal, Oct 31, 2026. A script played Muse against the real connector, on a laptop running LFM3-VL-2B, with invented records.

The essays stay on the laptop. The result waits until the teacher says yes, and the plan with real names is saved only there.

Tool Muse called
run_private_task
What came back
started
Time
0.1 s
Private values in the output
None
Step 1 of 5

What happens to a result

  1. Your device reads the files

    The documents and the local model stay on hardware you own. Muse never receives the files.

  2. Names become tags

    Exact checks and a name pass turn names and identifiers into tags like [Student A] before anything can leave.

  3. Your rules decide

    A policy file you control, not a model, says what can be shared, what waits for you and what never leaves.

  4. You approve what waits

    A card shows you the exact text on a page only your device serves. Muse hears that it's waiting for your OK.

  5. One last check, then tags

    A final check looks for anything private. If a check can't run, the result stays on the device.

What the tests show

0
leaks of names or identifiers across 300 invented items on the device
82.0%
agreed with the expected decision on 200 items (164 of 200), and 81.0% on 100 tasks
0 of 18
tool outputs held a private value in the rehearsal above

"Leaks" counts names and identifiers found in released text, and doesn't cover everything. 5 of those 300 results were shared when they should have waited or stayed on the device. The reports are in privacy/eval/reports in this project's repository.

Who it's for

  • Teachers: grading and feedback, the one job measured from start to finish.
  • Clinic front desks: counts, referral lists and reminders, with HIPAA and 42 CFR Part 2 toggles. Not measured as a full job yet.
  • Finance and tax offices: summaries over client files, with GLBA, PCI DSS and tax preparer toggles. Not measured yet.
  • Law offices and HR: summaries over client or employee files. Not measured yet.
  • City casework: the government toggles send nothing to Muse, so the caseworker asks on the device. Not measured yet.

What isn't done yet

  • No real users yet. Every person, essay and record used here is invented, and the teacher's answers are simulated.
  • The numbers above come from a development set. The sealed test set hasn't been scored, so there's no unseen result yet.
  • Liquid's LFM3-VL-2B does every model step on the device today. Meta's Muse Glimmer hasn't been measured on a device for this.
  • In the real Muse app it paired, listed its tools and ran one search. The full scenes have only run with a script playing Muse.
  • Tagging can miss things, such as a name written in lower case that the model also misses, and a summary can point at someone without naming them.
  • Whatever you type into Muse reaches Meta before the device sees it.
  • A privacy toggle doesn't make anyone legally covered. Your school, clinic, agency or company decides whether Muse may be used.
  • It runs on a laptop. The Jetson path is designed and tested in simulation, not run on a board in this setup.

Run it on a laptop

It needs llama.cpp and the LFM3-VL-2B model. From the repository's privacy folder:

uv sync
cp .env.example .env   # DEVICE_ONLY=1 is already set
scripts/start.sh

To add it to Muse, deploy the relay once, connect the device to it, add the custom connector in the Muse app and type the pairing code. The full steps are in privacy/docs/MUSE_SETUP.md.

An independent project, not made or endorsed by Meta. Muse Privacy shares no code with the rest of Muse Muse, and it isn't part of this site's deployment.