Brilliant Labs

Brilliant Labs is a chat with images tool. Open-source AR glasses platform with a microOLED display, an on-device NPU, a Lua scripting layer on ZephyrOS and a cloud AI companion called Noa that holds multimodal conversations about what the wearer sees and hears.

Brilliant Labs builds Halo, a pair of open-source smart glasses aimed at people who want to modify their own hardware rather than accept a sealed consumer product. The frame carries a color microOLED display, an optical sensor used for AI inference, two microphones with audio activity detection, a six axis IMU with tap detection, and dual bone conduction speakers. An Alif B1 processor pairing a Cortex-M55 CPU with an NPU runs lightweight models on the device itself, while Bluetooth 5.3 handles the link to a phone. Lenses ship in optical grade stock form with anti reflective coating, and prescription or sunglass options are offered. The software side is where the project separates itself. Halo runs ZephyrOS with a Lua interface, so owners can write or swap scripts that control what the glasses do. A cross-platform mobile app handles setup and a cloud-based AI agent carries the heavier conversational work. That agent is Noa, which holds realtime multimodal conversations about whatever the wearer sees, hears or imagines. Noa is described as remembering what it saw, heard and said, so its responses grow more personal over time and can reinforce the wearer's own recall. Demonstration videos show it answering questions about food, translating signs on a bike ride, coaching a workout and giving boxing pointers. The obvious audience is developers, students and hardware tinkerers who treat eyewear as a platform. Anyone expecting a polished assistant that works out of the box should look at mainstream AI glasses instead, since the appeal here is access to the firmware and the community around it. Brilliant Labs frames the whole thing as an open ecosystem for developers and creatives, with hardware and software both published openly. Battery life is quoted at 14 hours, unusually long for this category and likely tied to the low power chipset and the fact that heavy inference leans on the cloud agent. Creators can also use the display for lightweight visual experiments. The trade-off is maturity: this is a hacker-friendly device with a scripting layer, not a finished lifestyle product.