GitHub radar
Open source conversational AI system designed for phone agents
Pipecat introduces PhoneLLM — an open-weights model for voice agent applications with low latency, multi-turn dialogs, and accurate tool calling.
PhoneLLM Alpha 1 is a 32B open-weights model built by Pipecat for voice agent applications. Voice doesn’t work well with most language models because of its slow first token latency, multi-turn context loss, and inconsistent tool invocation during conversation. PhoneLLM overcomes these challenges with low latency, robust multi-turn dialog, and accurate tool invocation. It is said to be competitive with GPT Terra 5.6 while being 94% cheaper and 1.3 seconds faster (P95 time to first token). Licensed under the BSD-2 License.
Why a vibe-coder should care
With an open model, you don’t have to pay for each call to the cloud. Deploy your phone-answering bot, customer service bot, or appointment booking bot on your own server with lower latency and cost compared to bleeding edge models.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Look up pipecat-ai/phonellm-alpha-1 on Hugging Face (https://huggingface.co/pipecat-ai/phonellm-alpha-1), find how to run it locally via llama.cpp or vllm, and show a simple voice dialogue example with a tool call.
A Mac with 32 GB of RAM or more, or a GPU with 24 GB — the 4-bit version takes around 19 GB.
Open on Hugging Face▌ More finds