Meta open-sources Muse code for AI gadgets; builder demos Pi gaze tracker

Meta opens Muse to third-party hardware

Meta has open-sourced the code needed to build AI gadgets around its Muse model, according to a developer post on dev.to dated October 26, 2026, opening the door to third-party hardware projects. The post illustrates the release with a working example: a gaze-tracking device assembled from a Raspberry Pi to support at-home vision therapy for amblyopia.

The shape of the source matters here. The dev.to entry is a first-person build account rather than a Meta announcement, and every figure in it comes from the author's own measurements or the project's documentation. The central claim, that Meta has made the relevant code openly available, reflects the post's framing of the news and is not corroborated by Meta within the text provided.

What the post says about Muse

The author describes Muse as a lightweight transformer designed to run on-device, kept small enough for embedded hardware through quantization to INT8. According to the post, the resulting model occupies less than 4 MB, which allowed a Raspberry Pi to run inference at 60 frames per second without a cooling fan.

The post also cites a Muse research paper that proposes a contrastive objective for few-shot adaptation. After replacing the default training loss with that objective, the author reports that calibration error on the gaze tracker fell from 2.4 degrees to 0.7 degrees in under 200 epochs, using only a small set of labelled therapy sessions as training data.

How the build works

The project began by cloning the Amblyotube repository onto a spare Raspberry Pi. The repository's documentation, per the post, promises sub-30 millisecond frame-to-photon latency when the pipeline runs alongside Muse. The author attached an inexpensive infrared camera, adjusted the camera calibration settings, and watched a gaze heat-map appear in a terminal, jittery at first and then steady once the Kalman filter settled.

The author frames the payoff in human terms: a child laughed as an on-screen butterfly followed her eye movements without perceptible lag. In that account, the open-source project supplied the underlying infrastructure, Muse supplied the on-device intelligence, and together they turned a research prototype into a device a family could use at home.

What remains unverified

Several things are left open. The exact contents of Meta's release, its licensing terms and its official documentation are not described in the post, and none of the performance figures have been independently tested. The post also does not address clinical validation, a notable gap for a demonstration device used in the context of a medical condition. Anything health-adjacent built this way would need proper evaluation before being relied upon.

Why it matters

If the release matches the post's description, it puts capable on-device AI within reach of hobbyist hardware budgets. A model that fits in under 4 MB, runs at 60 frames per second on a fanless single-board computer and adapts to a new task in a few hundred training steps changes who can build AI hardware, extending it beyond large companies to individual developers with a spare Pi and a cheap sensor. The application sketched in the post, an assistive vision-therapy device, sits squarely in the category of low-cost, low-latency products where such releases matter most. The caveat is that the story currently rests on a single builder's account, so the details warrant confirmation before being treated as settled.