Turning a Bluetooth Speaker into a Smart Assistant with Qwen 3.5‑Omni

In a previous post the author introduced Qwen 3.5‑Omni, highlighting its native internet‑search capability, tool‑calling support, and audio‑output voice synthesis, which together enable a large‑model‑powered smart speaker.

Using the real‑time API of Qwen 3.5‑Omni together with a locally running voice‑wake‑up model, the author built a demo that runs on a continuously powered Mac mini. An external microphone captures speech, the wake word "小千小千" activates the system, and the model answers queries; when the answer requires up‑to‑date information, the model performs a live web search before responding.

The demo successfully turns a Bluetooth speaker into an interactive assistant, but the author observes noticeable latency, likely caused by network round‑trips to the API endpoint. They suggest that a fully local deployment of a sub‑10 B open‑source version of the model would eliminate this delay.

The accompanying code, originally shared by Vibe coding, runs as‑is; the author notes it could be further optimized and eventually released as open source. Prompt engineering details are pinned in the comment section of the original post.