Daily

Daily is the realtime voice, video, and AI infrastructure company behind Pipecat, the open source orchestration framework that developers use to build voice and multimodal AI agents, deployed on Daily's global infrastructure for production workloads . This launch introduces PhoneLLM, an open-weights model tuned specifically for the constraints of production voice agents, where response latency between conversational turns matters as much as raw accuracy. It targets teams building phone-based customer support, lead qualification, and similar agent workloads that need reliable tool calling and instruction following without the delay introduced by reasoning traces. The launch matters right now because most production voice agents are still running on older general-purpose models like GPT 4.1, since newer state-of-the-art models default to extended thinking before producing output tokens. PhoneLLM is positioned as an alternative in that gap, claiming performance comparable to a much larger frontier model on typical voice agent tasks while running at a fraction of the latency and cost. The model is available on Hugging Face and is described as the first in a planned series of open releases coming out of Daily's model training and evaluation work. The release is led by Kwindla Hultman Kramer, co-founder of Daily and creator of Pipecat , and reflects work the team has been doing with enterprises to adapt open weight models to specific voice use cases, including building the evaluation and benchmarking infrastructure needed to measure whether a model is actually improving on the metrics voice agents care about. For founders and operators already running voice agents in production, this is worth a look as a candidate for cost and latency reduction on existing pipelines.