JA EN
shiichan

Pipecat PhoneLLM Mini lands on Workers AI for voice agents!

Cloudflare Changelog ★★★★ View Markdown
#cloudflare#ai#serverless

Hi, it's Shii-chan! A model for phone-style realtime conversations just arrived on Workers AI.

Cloudflare Changelog developers.cloudflare.com

What was announced?

The Cloudflare Changelog says @cf/daily-co/phonellm-mini-alpha-2, a model from the Pipecat team at Daily, is now available on Workers AI. PhoneLLM is built for voice agents and other realtime AI apps, where the model has to reply quickly and call the right tool mid-conversation.

Why it matters

In a voice agent, the language model sits between speech-to-text and text-to-speech, so every second the model takes adds to the caller's wait.

  • A fast model without reasoning mode can miss instructions or tool calls
  • A slower model in reasoning mode handles both well, but takes longer

Developers usually had to choose between the two.

What changes

Pipecat post-trained NVIDIA Nemotron 3 Nano 30B on phone and customer support scenarios, so PhoneLLM follows instructions and calls tools better without reasoning mode. You may not have to trade speed for accuracy as often!

Dive Deep

You can use it through:

  • The Workers AI binding (env.AI.run())
  • The REST API at /ai/run

Pipecat's recommended settings:

  • temperature: 0
  • Also set chat_template_kwargs: { enable_thinking: false }

Pipecat is an open-source Python framework for voice agents, and since Python Workers are generally available, you can run it on Cloudflare. The example:

  • Runs in a Durable Object
  • Uses WebRTC as the voice transport through the Realtime SFU
  • Uses DeepSeek V4 Flash by default

The changelog includes a prompt you can give your coding agent to deploy it with PhoneLLM. See the PhoneLLM Mini model page and Pipecat's model card for more.

Wrap-up

  • PhoneLLM Mini is available on Workers AI
  • Based on Nemotron 3 Nano 30B, post-trained for phone and support scenarios
  • Better instruction following and tool calling without reasoning mode
  • temperature: 0 and enable_thinking: false recommended
  • Run Pipecat voice agents with Python Workers, Durable Objects, and WebRTC

Great if you're building low-latency voice agents!

Related posts