Meta Built the Wearable-First Model
Muse Spark matches Llama 4 Maverick at 10x less compute and ships on $299 glasses.
Meta did not shrink Llama to fit on glasses. They built a new model from scratch.
What Shipped
On June 23, Meta launched Meta Glasses - three self-branded frame styles starting at $299, available in 17 countries. No Ray-Ban logo. These are Meta's own.
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. What is not familiar is the AI running on them.
Muse Spark is the first model from Meta Superintelligence Labs. It is not a distilled Llama. It is a ground-up rebuild - 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.
The model is natively multimodal. Text, images, and voice are processed simultaneously rather than patched together through adapters. Its Contemplating mode runs multiple agent instances in parallel - generating solutions, self-refining, and aggregating into a final output. Multi-agent reasoning at inference time, running on a chip in your glasses frame.
The Pattern
For years, deploying AI on edge devices meant one thing: take your cloud model, distill it, quantize it, and hope it still works. Muse Spark breaks that pattern. Meta started with the constraint - battery power, thermal limits, latency requirements of a wearable - and built the model architecture around it.
This is the difference between fitting a model into a device and building a model for a device.
There is a strategic signal in the licensing too. Llama is open-source. Muse Spark is not. Meta open-sourced the model that runs everyone's cloud infrastructure but kept the one that runs on their glasses. The wearable model is proprietary because the wearable is the product.
What Changes for Builders
If you build ambient AI systems - and I do - the Muse approach changes how you think about model selection for wearables. The old calculus was: pick the biggest model you can afford to distill. The new calculus is: purpose-build for the form factor.
10x compute reduction while matching the reference model is not incremental. It is a different design point. It means the model can reason through complex queries on a battery budget that lasts a full day.
For the wearable agent ecosystem that Qualcomm just commoditized with START, Muse Spark is what runs on top. The hardware is a commodity. The model is the moat.
Where This Heads
Muse Spark is the first in a scaling series. Each generation validates before going bigger. The architecture was born wearable-first. That origin point shapes everything that follows.
The companies that win wearable AI will not be the ones who shrink their cloud models the fastest. They will be the ones who never started from the cloud.