I've been experimenting with building low-latency voice agents that run fully on-device (Apple Silicon, MLX) using Kyutai's Moshi, an open-weight, full-duplex speech model. It's genuinely fast and natural to talk to. But I've hit a wall and I'd love help if anyone has cracked it. The problem: tool calling / function calling with full-duplex models. Moshi has no tool-calling head. It's a speech-to-speech dialogue model, so there's nowhere to plug a tool result in. My setup: an LLM (via OpenRouter) detects the intent and calls the tool (an orders DB lookup), and the answer has to be spoken by Moshi, in its own voice, not a second TTS voice that would jar the user. What I tried: Feed the tool result back into Moshi as input audio ("say this back: …"). I ran a 56-trial test. Only ~7% came back faithful. Moshi treats it as conversation ("Thanks, I appreciate it!") instead of repeating, and once it turned "shipped" into "placed". Not usable. Force Moshi's text tokens by swapping in a custom text sampler so Moshi has to speak my exact words. This is far better: 79% overall, and short sentences hit ~94% on key facts / ~78% word-for-word. It only works with lower audio-sampling temperature and word timing copied from Moshi's own speech. Lists and long sentences still truncate, and it sometimes slips ("shipped" becomes "dipped"). So I'm stuck between "too unreliable" and "fine-tune a 7B model on a rented GPU." Questions for anyone building voice AI agents with open-weight models (Moshi, Sesame, Qwen-Omni, MiniCPM-o, etc.): Has anyone shipped tool calling on a full-duplex model? Cascade (STT, LLM, TTS), text-forcing, fine-tuning, or a hybrid? How do you get a model like this to say an exact sentence at ~98% accuracy? Any papers, repos, or people I should read or follow? Happy to open-source my test harness and share results. Replies, RTs, and pointers all welcome.