Muse Gadgets Open Source: Meta Opens the Hardware, Not the AI

Meta has released Muse Gadgets, an open-source toolkit for building your own hardware around its Muse AI assistant. The code is free, permissively licensed and aimed squarely at hobbyists. But the openness has a clear boundary. What Meta opened is the plumbing that connects a homemade device to Muse. The intelligence behind it remains Meta's. Given the company's recent history with open-source AI, that distinction matters more than usual.

What Meta actually released

Muse Gadgets consists of firmware and software development kits published on GitHub in the facebookincubator/muse-gadget-sdk repository under the Apache 2.0 license 12. It comes in two parts. An ESP32 Device SDK targets cheap microcontroller boards, and a Linux Device SDK turns a Raspberry Pi or another Linux machine into a Muse device 12.

The workflow is simple. You flash the firmware onto an off-the-shelf board, pair it with the Muse app on your phone, and Muse can then use whatever is wired up: speakers, microphones, small screens, buttons, sensors and actuators 12. Developers get started by obtaining an SDK token from gadgets.muse.ai 2. Meta's own pitch describes Muse gadgets as "open-source devices you build yourself." Its suggested projects include a color E Ink display that shows reminders and an HDMI stick that puts the agent on a bigger screen 5. One enthusiast write-up frames the appeal as turning a roughly $15 circuit board into a talking AI companion, with weekend builds like a desk "puck," a watch and an ambient e-ink display 1.

According to one account, the launch came from Meta Superintelligence Labs on October 2, 2026. Nat Friedman, the lab's head of product, presented it as a side project, and Chief AI Officer Alexandr Wang promoted it as the start of a "Muse ecosystem" 2. That report also says a free USB-C smart-home hub called Muse Home Link arrived alongside it 2. Separately, Meta appears to be exploring its own Muse hardware, including a Tamagotchi-like device called Muse Charm about which little is known 5.

Where the 'open' ends

The hands-on guide spells out the key caveat directly: you are not installing Muse itself on the board 1. The device works as a peripheral. It handles input and output and relays them through the paired phone app to an assistant that runs elsewhere. Anyone can inspect, modify and redistribute the Apache 2.0 firmware. The model that does the reasoning is not part of the package.

This would be a technicality if Meta had not changed its stance on model openness. Earlier this year it effectively moved on from its open-weight Llama family in favor of Muse Spark, a proprietary model built by Meta Superintelligence Labs 4. The New Stack argued the shift was driven by profit. It noted that Meta's promise that current Llama models "will continue to be available as open source" covers existing releases only and says nothing about future ones 4. The outlet also pointed out that Llama's "open source" label was disputed in the first place 4.

The Glimmer complication

The picture is more mixed than a simple retreat from openness, though. VentureBeat reports that Meta has released Muse Glimmer, a 30-billion-parameter open-weight model tuned for running autonomous agents on high-end consumer Macs and PCs 3. It ships under Apache 2.0, which is more permissive than Llama's custom license with its 700-million-monthly-user threshold, and the weights are on Hugging Face 3. VentureBeat calls it Meta's first fully open release since Muse Spark replaced Llama in April, with support announced for Ollama, LM Studio, vLLM, llama.cpp, MLX and other runtimes 3.

So Meta now runs two tracks. Its flagship assistant is closed. A smaller agent model is genuinely open-weight. And a hardware SDK is open but points at the closed service. None of the reporting on Muse Gadgets says that gadgets can run on Glimmer locally instead of the hosted Muse. Builders should not assume they can until Meta says so.

Reading the strategy

The most plausible reading is that Muse Gadgets is an ecosystem play, not a return to open AI. Opening the device layer costs Meta little. It invites makers to put Muse on desks, wrists and walls, and every one of those devices feeds the proprietary assistant. Wang's "ecosystem" framing 2 and Meta's interest in its own Muse hardware 5 fit that interpretation. Under this model, open firmware widens the reach of a closed product.

That isn't necessarily a bad deal for hobbyists. Apache 2.0 firmware for ESP32 and Raspberry Pi is useful, and cheap AI-connected builds are fun. But "open source" here describes the edges of the system, not its core. Developers who left Llama over openness concerns 4 will find Glimmer 3 far more relevant than Muse Gadgets. Anyone building a gadget should remember that its brain still depends on Meta's servers and Meta's terms.