Cloudflare adds PhoneLLM Mini, a model built for voice agents
Daily built PhoneLLM Mini for low-latency phone and support calls; Cloudflare now serves it on Workers AI with tool calling and a 262,140-token context window.
Cloudflare added PhoneLLM Mini Alpha 2 to Workers AI on 15 October. The model, built by the Pipecat team at Daily, is a fine-tune of Nvidia’s Nemotron 3 Nano 30B-A3B aimed squarely at voice agents: phone support, IVR replacement, anything where a model has to reply fast and call the right tool mid-call.
It runs with thinking disabled by default, trading reasoning depth for speed, and ships with a 262,140-token context window and function calling. Cloudflare prices it at $0.06 per million input tokens and $0.24 per million output tokens, and serves it through the usual env.AI.run() binding, the REST API, or an OpenAI-compatible /v1/chat/completions endpoint.
const response = await env.AI.run('@cf/daily-co/phonellm-mini-alpha-2', { messages });Why it matters: voice agents live or die on latency, not cleverness. A model tuned to skip reasoning and answer fast, running on the same platform as Cloudflare’s realtime audio stack and Python Workers, removes a round trip to a separate frontier API for the part of the pipeline that most needs to be quick.
The caveat: it’s labelled Alpha 2 and narrowly built for phone-style conversations. It is not a general-purpose chat model, and Cloudflare gives no indication of when, or whether, it graduates out of alpha.