Cloudflare has added the @cf/daily-co/phonellm-mini-alpha-2 model to the Workers AI catalog. The model is positioned for real‑time voice agents where each millisecond of inference adds to caller wait time, and it has been post‑trained to handle instruction following and tool calls without needing the higher‑latency reasoning mode.
New model and inference profile
PhoneLLM Mini is derived from an NVIDIA Nemotron 3 Nano 30B base that was further trained on phone‑call and customer‑support data. The training objective lets the model produce correct tool‑call JSON and obey prompts while staying in the fast, non‑reasoning execution path. This contrasts with the typical trade‑off where a fast model may miss tool calls and a slower model with reasoning mode handles them reliably.
Calling the model from Workers
Developers can invoke the model using either the Workers AI binding or the REST endpoint. The binding call looks like:
env.AI.run("@cf/daily-co/phonellm-mini-alpha-2", {
temperature: 0,
chat_template_kwargs: { enable_thinking: false }
});Via HTTP the request is posted to /ai/run with the same payload structure. Pipecat’s guidance recommends a temperature of 0 and disables the internal "thinking" flag to keep latency minimal.
Running a Pipecat voice agent on Cloudflare
Pipecat is an open‑source Python framework for building voice agents. With Python Workers now generally available, a Pipecat agent can be deployed inside a Cloudflare Durable Object. The reference implementation uses WebRTC through the Realtime SFU for bidirectional audio transport and defaults to the DeepSeek V4 Flash model. Swapping the default model for PhoneLLM Mini only requires updating the model identifier and the chat_template_kwargs as shown above.
Operational considerations
- Latency vs. correctness: PhoneLLM Mini aims to keep inference time low while still handling tool calls, reducing the need to toggle reasoning mode.
- Configuration defaults: Use
temperature: 0andenable_thinking: falseas recommended; deviating may re‑introduce latency or unpredictable output. - Durable Object scaling: The voice‑agent example runs in a Durable Object, which provides automatic state sharding and failover. Operators should monitor object instance counts and memory usage as call volume grows.
- WebRTC and SFU usage: Real‑time audio is carried over the Realtime SFU. Ensure the SFU is provisioned for the expected concurrent streams and that network latency is within acceptable bounds for voice quality.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Teams building voice‑driven AI services can now replace a slower reasoning‑mode model with PhoneLLM Mini to shave milliseconds off each turn without sacrificing tool‑call reliability. The migration path is a single model identifier change in the Workers AI call and a tweak to the chat template flags. Practitioners should benchmark end‑to‑end latency after the switch, verify that tool‑call JSON is still produced as expected, and monitor Durable Object resource usage as call volume scales.
