Related project: AI agent on Galaxy Watch with 2.8 MB binary

[ThinkOffApp]

Your work on fast multimodal LLM inference on mobile devices is relevant to what we're doing with ClawWatch (https://github.com/ThinkOffApp/ClawWatch).

ClawWatch runs an OpenClaw AI agent natively on Galaxy Watch. The on-device runtime is NullClaw, a 2.8 MB static ARM binary written in Zig. Voice input uses Vosk for offline speech recognition. LLM inference currently goes through a network gateway rather than running on-device, which is the main limitation.

If mllm's inference engine could run small models directly on watch-class hardware (Exynos W930/W1000, 2 GB RAM), that would enable fully offline AI agents on wearables. The agent runtime and voice pipeline already work locally. The missing piece is local inference.

Would be interesting to know if anyone has tested mllm on Wear OS / watch-class ARM SoCs.

[chenghuaWang]

On macOS, the runtime binary size to run an LLM is 5.9MB. However, the hardware limits of chips like the Exynos W930/W1000 with only 2GB of RAM are so tight that even the Qwen 0.6B model might not be able to run. We will likely need even smaller models. I'm hoping someone in the community can provide some insights or pointers on this.