cd ~/
Wearable AIOn-Device AIAgentic AIEdge ComputingModel Architecture

Muse Spark Starts From the Battery

Muse Spark matches Llama 4 Maverick at 10x less compute and ships on $299 glasses.

Evyatar Bluzer
3 min read

On June 23, Meta launched Meta Glasses - three self-branded frame styles starting at $299, available in 17 countries, with no Ray-Ban logo in sight. The hardware is solid but familiar: 12MP ultrawide camera, 3K video, five-microphone array, eight-hour battery with a charging case that holds another 40 hours. I want to talk about the model instead, because Meta did something there that I have been waiting years for someone to do.

Three weeks ago I passed along Meta's plan to start shipping new frames in June. It kept the date.

What is Muse Spark?

Muse Spark is the first model from Meta Superintelligence Labs, and it is a ground-up rebuild rather than a distilled Llama: new architecture, new data pipelines, new training infrastructure built on Meta's Hyperion data center. The key number: Muse Spark matches Llama 4 Maverick performance at 10x less compute. It is natively multimodal, processing text, images, and voice together instead of patching them in through adapters, and its Contemplating mode runs multiple agent instances in parallel - generating solutions, self-refining, aggregating into a final output. Multi-agent reasoning at inference time, on a chip in a glasses frame.

The Pattern That Broke

For years, deploying AI on edge devices meant one thing: take your cloud model, distill it, quantize it, and hope it still works. Meta started from the other end. The constraints of a wearable - battery power, thermal limits, latency - came first, and the model architecture was built around them. You can fit a model into a device or you can build a model for a device, and until this launch only the first path had shipped at scale.

Two paths to a wearable modelTwo rows of boxes joined by arrows from left to right: the top row runs from a cloud model through distill and quantize to the wearable; the bottom row runs from wearable constraints through architecture to Muse Spark and Meta Glasses, with its arrows highlighted. THE PATTERN THAT BROKE: FIT A MODEL INTO A DEVICE Cloud model distill quantize hope it still works Wearable MUSE SPARK: BUILD A MODEL FOR A DEVICE Wearable constraints battery, thermal, latency Architecture built around them Muse Spark natively multimodal Meta Glasses from $299 matches Llama 4 Maverick at 10x less compute
Two ways to get a model onto a wearable: the old path distills a cloud model down until it fits, while Muse Spark started from the frame's battery, thermal, and latency limits and matches Llama 4 Maverick at 10x less compute, per Meta.

The licensing carries a strategic signal too. Llama is open-source; Muse Spark is not. Meta open-sourced the model that runs everyone's cloud infrastructure and kept the one that runs on their glasses. The wearable model is proprietary because the wearable is the product.

The New Calculus

If you build ambient AI systems - and I do - Muse changes how you pick models for wearables. The old calculus was to grab the biggest model you could afford to distill. The new one is to purpose-build for the form factor. A 10x compute reduction while matching the reference model is a different design point altogether, one where the device can reason through complex queries on a battery budget that lasts a full day. For the wearable agent ecosystem Qualcomm just commoditized with START, Muse Spark is what runs on top. The hardware is a commodity. The model is the moat.

The Scaling Series

Muse Spark is the first entry in a scaling series - each generation validates before the next goes bigger - and the architecture was born wearable-first, which shapes everything downstream of it. My verdict: the winners in wearable AI will be the teams that never started from the cloud, and anyone still distilling is now a visible generation behind.

Comments