RunanywhereAI/RCLI

Implements a complete STT + LLM + TTS + VLM pipeline optimized for Apple Silicon via MetalRT, a proprietary GPU inference engine delivering sub-200ms latency and 714x faster speech recognition than real-time. Includes hybrid vector+BM25 RAG (~4ms retrieval), 40 native macOS voice actions (Spotify control, app launching, system commands), and on-device vision capabilities for camera/screen analysis without model quantization. Supports hot-swappable models (Qwen3, LFM2, Whisper) with graceful fallback to llama.cpp for M1/M2 Macs.

1,054 stars. Actively maintained with 89 commits in the last 30 days.