Voice agents no longer trade speed for accuracy: PhoneLLM Mini lands on Workers AI, following instructions and calling tools with thinking off
Synopsis
PhoneLLM Mini, a model from the Pipecat team at Daily, is now available on Cloudflare Workers AI for voice agents and other realtime AI applications. In a voice agent the language model sits between speech-to-text and text-to-speech, so every second it takes adds to the caller's wait; developers usually choose between a fast model without reasoning mode, which can miss instructions or tool calls, and a slower model in reasoning mode, which handles both well. 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.
Interpretation
Developers building voice agents can now call a model tuned for phone and customer support scenarios directly on Cloudflare Workers AI, instead of choosing between a fast model that misses instructions and a slow one that does not. The prior choice was between a fast model without reasoning mode, which can miss instructions or tool calls, and a slower model in reasoning mode, which handles both well but adds to the caller's wait. 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.
Access follows the two existing Workers AI paths, with the model offered under the identifier @cf/daily-co/phonellm-mini-alpha-2. Calls go either through the Workers AI binding env.AI.run() or through the REST API at /ai/run, matching how other models on the platform are used. Pipecat recommends temperature: 0 and setting chat_template_kwargs: { enable_thinking: false }, which turns thinking mode off.
Perspective
This item addresses developers building voice agents and other realtime AI applications, where a model has to reply quickly and call the right tool mid-conversation. The model is offered on Workers AI under the identifier @cf/daily-co/phonellm-mini-alpha-2 and can be called through the binding env.AI.run() or the REST API at /ai/run, with Pipecat recommending temperature: 0 and chat_template_kwargs: { enable_thinking: false }. The material also points to a Pipecat voice agent example that runs in a Durable Object and uses WebRTC as the voice transport through the Realtime SFU, defaulting to DeepSeek V4 Flash and deployable with PhoneLLM.
The material gives no comparison of PhoneLLM against fast models or reasoning-mode models on response latency, instruction following, or tool calling, so the statement that it follows instructions and calls tools better without reasoning mode describes the post-training target relative to its base model rather than a quantified comparison with other models. The model identifier carries an alpha suffix, and the material does not state a service level or long-term availability.