Muse on devices and robots: edge boards, butlers and glasses

Muse shows up on hardware in two ways. Muse Glimmer 30B, the open-weights model, runs locally on edge boards, laptops and desktop GPUs. Muse Spark 1.3, the cloud model, acts as the brain for devices that call the Meta Model API. On the consumer side, Meta has announced Muse on AI glasses and a pocket device called Muse Charm, but neither has full details yet.

The catalog's robotics and devices section is small. Here is everything in it, plus the local-hardware numbers people reported.

Can Muse Glimmer run on an edge device?

Yes, slowly. Samuel Alexander ran Glimmer 30B on a Qualcomm Dragonwing IQ-9075 board for zero-shot PCB defect inspection and tool calling. He measured 21.6GB resident with the full 131K context and 2.84 tokens per second generation. That is usable for inspection jobs where a verdict can take a minute, not for conversation.

NVIDIA's technical blog covers Glimmer on Jetson, alongside RTX 5090, DGX Spark and DGX Station (NVIDIA). Meta lists ExecuTorch and Arm among launch partners (Meta).

What hardware does Muse Glimmer need?

Meta says Glimmer fits under 20GB at 4-bit and targets 24–32GB devices such as M4/M5 Max Macs and the RTX 5090. Builders pushed lower and higher, with results as reported:

Hardware	Reported result
RTX 4060 laptop, 8GB VRAM	Q4_K_M via llama.cpp for bug fixing and invoice extraction
12GB laptop GPU	EXL3 3.00bpw at 100K context, ~30 tok/s
Single RTX 4090	130K context in 19.3GB VRAM, 3,100+ t/s prefill
DGX Spark / GX10	29.6 tok/s with DFlash, up from 11.9
Dell Pro Max GB10	Strong vision scores but failed five deployment gates
M5 Pro via oMLX	DFlash speedup only about 1%

Two notes. The Dell GB10 report is a useful negative result and worth reading before you buy. And DFlash speedups vary a lot by setup: Meta's model card reports 1.8x on an M5 Max, while the M5 Pro test above saw almost none.

For Apple Silicon specifically, see muser, which runs on Metal with an optional lane where an NVIDIA GB10 node prefills and hands the KV cache to the Mac.

Is there a robot running on Muse?

Two entries.

Tarquin, a StackChan robot butler. Custom firmware and a Go server turn an M5Stack StackChan into a voice-driven butler with wake word, face tracking and on-device speech, using Muse Spark 1.3 via the Meta Model API as its brain. This is the most complete hobby robot in the catalog, and the pattern is common: cheap local hardware for sensing and speech, a cloud model for reasoning.

Muse Spark 1.2 robot navigation. Meta's own demo of Spark reading multimodal observations and calling tools to guide a robot through an unstructured environment. It is a demo, with no public code.

Can Muse control smart home devices?

Meta demoed Glimmer building a Home Assistant dashboard: it found a local Home Assistant instance through network tool calls, queried device APIs, wrote an HTML/CSS/JS dashboard and served it locally to check it worked. Because Glimmer runs locally, nothing leaves the network.

On the personal agent, one user probed Muse's tool schemas for about 90 minutes and reported HomeKit control, geofences and BLE scanning absent from public docs. That is one person's self-scan, not a Meta feature announcement. A separate privacy teardown documents shipped tools for screen control and background sync.

The homelab control plane build joins the Muse VM to Tailscale so the agent can run commands on a home Ubuntu server, and Tailscale is a launch connector.

Can Muse make phone calls?

Community tools add this. One MCP provisions a number for inbound and outbound calls to US and Canadian numbers, and another service gives Muse a real non-VoIP number with a mode where it whispers reminders mid-call. Matt Schlicht's Muse called five Tesla stores at once.

What about Muse on glasses and Muse Charm?

At Connect 2026 Meta announced (Meta):

• Glasses integration. Muse on AI glasses, described at launch as coming "in the coming months."
• Muse Charm. A pocket device with a real-time voice model. Meta says details come "later this year" (Muse Charm).
• Voice mode and a realtime avatar for the existing apps.

We have no hands-on reports for either device yet, so we will not describe how they work.

Local model or cloud model for a device?

The choice comes down to memory, latency and privacy. Glimmer keeps data on the device but needs roughly 20GB and runs at single-digit tokens per second on edge boards. Spark needs a network connection and an API key, costs per token, and gives you a 1M-token context and max reasoning. Tarquin splits the difference: wake word, speech and face tracking run locally, and only the thinking goes to the cloud.

Where to start with Muse on hardware

1. If you want offline and have 24GB or more, start with Glimmer in Ollama or LM Studio.
2. If your device is small, keep the model in the cloud and call Spark over the API like Tarquin does.
3. Measure on your own board before trusting any tok/s figure, including ours.

Browse robotics and devices, and see Muse for developers for the API side.

Frequently asked questions

Can Muse Glimmer run on a Raspberry Pi-class board?

The closest catalog report is a Qualcomm Dragonwing IQ-9075, which needed 21.6GB resident and generated 2.84 tokens per second. Boards with much less memory are unlikely to fit the 30B model without heavy quantization.

Can I use Muse as the brain of a robot?

Yes. The Tarquin StackChan build uses Muse Spark 1.3 through the Meta Model API for reasoning, with speech and face tracking on the device. Meta has also demoed Spark 1.2 guiding a robot.

When is Muse coming to Meta glasses?

Meta announced glasses integration at Connect 2026 and earlier said it would arrive in the coming months. No exact date was given.

What is Muse Charm?

A pocket device with a real-time voice model, announced at Connect 2026. Meta says more details come later this year.