Offline Conversational LLM on Pi

I’m turning a plush toy into a bilingual chatterbox that never needs Wi-Fi. Inside the toy sits a Raspberry Pi 5 with 8 GB of RAM; your task is to pull an open-source large-language model (roughly 8 GB–24 GB) from Hugging Face, streamline it, and make it run locally on that board. The goal is simple spoken conversations in both English and Spanish, so the full loop—speech-to-text, prompt handling, text-to-speech—has to live on the Pi.

I already have a fresh 64-bit Raspberry Pi OS install, a USB microphone/speaker HAT, and plenty of storage. What I need from you is: the model selection, any quantisation or pruning that keeps RAM usage within limits, all required runtime libraries (llama.cpp / GGUF, Whisper.cpp, or whatever toolchain you prefer), and clean startup scripts that boot straight into conversation mode when the toy is powered on. Latency should feel natural enough that a child can ask a question and hear a response without noticeable lag.

Deliverables
• The optimised model files or exact download links
• Installation and build scripts with inline comments
• A single launch script or systemd service that autostarts the conversation loop
• Brief README explaining how to switch between English and Spanish during use

I’ll flash the image you provide, power the toy, and run a quick two-minute demo in both languages; if it holds the conversation without network access, the job is done.