Coming soon? Meta smart glasses are equipped with "facial recognition" feature: three AI models + biometric database

A security researcher disassembled the Meta smart glasses mobile app and discovered: it already has a complete on-device facial recognition pipeline built-in, including three AI models, a local biometric database, a cosine similarity vector index, and a "Person recognized" notification. This system can be run manually, but Meta has not yet enabled it for consumers.
(Background: Amazon, Meta, and the FBI can access an unpublicized intelligence-sharing network: How Seattle Shield forms the AI surveillance era)
(Additional context: Google partners with Samsung to launch new smart glasses: Audio glasses coming this fall to challenge Meta’s over 70% market share)

Table of Contents

Toggle

• Three models, one pipeline, 96 MB facial feature engine
• Unnamed face: local database and 0700 permission storage directory
• Launch expected by the end of the year

A complete machine capable of recognizing strangers has been assembled and is lying within the Meta smart glasses mobile app: three AI models, about 100 MB, a facial feature vector of 2048 numbers, plus a "Person recognized face recognition" notification. This is the latest discovery after a security researcher disassembled it.

The app supporting Meta smart glasses is called Stella; the researcher’s disassembled version is v273.0.0.21. What he found was not just a few lines of test code, but a fully connected, manually operable facial recognition pipeline—from face detection, alignment, feature generation, to local database query, and triggering high-priority system notifications, all seamlessly connected.

Three models, one pipeline, 96 MB facial feature engine

Stella app contains three ExecuTorch models, downloaded from Meta’s NMLML asset delivery system from Meta servers, totaling about 100 MB. Each model has its own role, arranged sequentially into a pipeline:

The first is SCRFD.pte (3.4 MB): responsible for detecting faces in the image.

The second is KPSAligner.pte (117 KB): crops detected faces, corrects angles, and aligns each face to a standard reference.

The third, and largest, is SFace.pte (96 MB): compresses an aligned face into 2048 floating-point numbers, which constitute the face’s "faceprint" or biometric fingerprint. Simply put, these 2048 numbers are the face’s coordinates in a mathematical space; two photos of the same person taken under different lighting or angles will produce vectors close in space, while different people’s vectors will be far apart.

The researcher tested this pipeline: directly calling the app’s existing handler, feeding in a publicly available photo of philosopher Michel Foucault. The entire pipeline ran successfully—detecting the face, generating the 2048-dimensional faceprint, querying the local vector index, and then triggering an Android high-priority notification titled "Person recognized," with the content "Recognized Michel Foucault."

Unnamed face: local database and 0700 permission storage directory

The app creates an SQLite database on the device, located under Meta’s cross-device synchronization framework RLDrive, with a namespace called person_profiles, designed to be filled remotely (from the server side). In simple terms, this database is not just local; its architecture reserves an interface for pushing contact data from Meta’s cloud.

The database structure includes a person table, a face table, and a virtual vector table. The recognition logic is: perform a cosine similarity search on this vector index—meaning, find the "closest" face coordinate in the database to the new face’s coordinate—and then join back to person.name to display as notification text.

Launch expected by the end of the year

Meta holds over 70% market share in the smart glasses market, and The New York Times, citing internal documents in early February 2026, reports that Meta is considering adding facial recognition features to its smart glasses, aiming for a launch by year’s end. TechCrunch also published related reports. This reverse engineering investigation provides concrete technical evidence supporting those external reports.

However, the researcher also clarified that currently, on the original device and with unactivated accounts, these user-facing UI elements do not appear, and the features are not yet enabled. He emphasized: "This is not Meta secretly recognizing the people you see. Instead: the entire device capable of doing that is already assembled, operational, and lying on your device, under Meta’s control."

For now, when or whether to activate this feature is Meta’s decision. The only requirement is to flip a switch—no additional development needed.