NVIDIA's PersonaPlex 7B model now runs natively on Apple Silicon via MLX, enabling full-duplex speech-to-speech processing in Swift. This breakthrough eliminates traditional transcription steps, processing audio directly at ~68ms/step with a real-time factor (RTF) of 0.87 — faster than human listening speed.

The 4-bit quantized model (~5.3 GB) leverages NVIDIA's PersonaPlex architecture adapted for Apple's unified memory and Metal acceleration. Key innovations include the Depformer component, which uses per-step weight switching to reduce model size by 3.7x without quality loss. The system reuses the Mimi audio codec from prior projects, ensuring seamless integration with existing speech processing pipelines.

Performance tests on M2 Max show PersonaPlex handles 17 parallel audio streams at 12.5Hz, maintaining sub-80ms latency per frame. The library supports end-to-end testing by round-tripping speech through ASR and back, validating contextual accuracy. Streaming capabilities now allow continuous audio chunk generation via `respondStream()`.

This development positions PersonaPlex 7B as a critical tool for building conversational AI on Apple devices, bypassing cloud dependencies. The open-source release at aufklarer/PersonaPlex-7B-MLX-4bit includes system prompts for role-based interactions, from customer service to technical support.