I'm a firmware/FPGA engineer working on a smart speaker/hub that runs a large open MoE model (Qwen-Flash class, not a 3B) entirely on the device, ≤30 W, no cloud, no subscription, Matter/Thread/Zigbee built in, works with HA. Simulations on real model routing show ~25-40 tokens/s by streaming experts from flash instead of needing 80 GB of RAM. Before building hardware I want to know what actually matters to you:

• Is a small local model (e.g. qwen3 4-8B on a mini PC) already good enough for your voice control?
• What would make you pay for a dedicated box instead of DIY?

• What price feels fair? 5-min survey link.

  Waitlist/info Link here