Local-Ai

A Qwen3.5 9B model, an RTX 3060 and a ridiculous travel itinerary. The local assistant finds real flights, then correctly admits it cannot make the day work.

ONNX quantisation tuning produced great RTF numbers and unusable audio. A GGUF runtime with worse numbers won on the only test that mattered.

How Fulloch’s CPU tier runs speech recognition and text-to-speech on a CPU-only PC

Same conversation, same prompt, same tools, four model sizes. Watching Fulloch degrade from a capable agent into something that confuses the World Cup with Formula 1.

A first year development diary of Fulloch, the Fully Local Home voice assistant.