phonellm-alpha-1: Lightweight LLM for Voice Agents Based on Nemotron

Fine-tuned NVIDIA Nemotron-3-Nano-30B-A3B to optimize for voice agents and phone call scenarios. By adopting a Mixture-of-Experts architecture, it operates at a 3B parameter level during actual inference despite its 30B parameter scale, enabling real-time conversation processing with low latency.

Designed for tight integration with the pipecat framework, it simplifies the pipeline for converting voice input to text and generating responses instantly. With support for function calling and tool use, the model can directly execute complex business logic such as reservation confirmations and information lookups.

Specialized for building conversational AI in English environments, it is free for commercial use under the BSD-2-Clause license. It is suitable for developers seeking to reduce latency issues and token waste common in voice conversations with existing general-purpose LLMs, and to implement natural voice interactions using lightweight resources.