A Full-Duplex Voice Assistant on One DGX Spark
@jayleaton’s LocalRouter runs Qwen 27B, speech, images, video and Whisper on TensorFold, and talks back in about half a second.
“Jarvis in a box,” its builder says.
@jayleaton’s LocalRouter AI assistant runs on one DGX Spark, full duplex, built on @ashxhart’s TensorFold. The stack: Qwen 27B, Drama 3 with streaming (he thanks @plotarmordev), Qwen Image Turbo 2.1, MiniMax H3 video and Whisper STT.
About 0.5 seconds after he stops talking, it starts with its first word. You can interrupt it anytime and set the response delay as you please.
Models make room for each other.
Normal models stay loaded and can be swapped as needed. Want a video model up? Then Whisper and Drama make space, and voice starts back up when the video is finished. He calls it a super early version of a fully modular design, MCP by default, with a simple UI you can extend. He asks whether people want TP=2 or Mac Studio versions next.
@MiaAI_lab calls it an insanely good package that runs on one DGX Spark.
Sources
- @MiaAI_lab on X · Oct 22, 10:22 AM PT
- @jayleaton on X ·