October 2026 Summaries

Daily is joining Deepgram, with Deepgram taking responsibility for selling, supporting, and developing Daily’s existing realtime audio, video, WebRTC, telephony, and developer infrastructure products while the Pipecat framework remains open source and vendor-neutral. An independent Pipecat Foundation governed by a steering committee representing the broader realtime AI ecosystem, including Deepgram competitors, will oversee the project’s future. The companies’ relationship dates to their shared participation in Y Combinator in 2016 and a transcription partnership beginning in 2019, and the combination reflects their shared focus on low-latency voice AI, infrastructure, models, and open-source orchestration. Daily describes Pipecat, open sourced in 2024, as a widely used framework for voice agents and multimodal AI that supports model switching, streaming, transport, turn-taking, interruption handling, context management, tool calls, observability, and other production requirements; it now has more than 200 integrations and nearly 400 contributors. The move is presented as an opportunity to expand work across the AI stack, use Deepgram’s engineering resources and voice technology, and develop open standards and tools for human-in-the-loop software as voice-based AI applications grow across customer service, devices, education, robotics, and other sectors.

Pipecat has released Alpha 2 of PhoneLLM, an open-weights text model optimized for low-latency voice-agent pipelines, with PhoneLLM Mini offering a 30B-parameter mixture-of-experts design for speed and compact deployment and PhoneLLM Max providing a 120B-parameter alternative for stronger accuracy. The update expands training data, improves performance across prompting styles, adds Pipecat Flows compatibility, and introduces “microthinking” in Max, which permits a single sentence of reasoning to improve tool selection without materially increasing latency. PhoneBench has also been broadened to assess error correction, instruction sequencing, tool-call errors, and structured conversation scenarios. The models can be self-hosted or deployed through AWS, Baseten, Cloudflare, and Modal, with reported high-throughput performance on NVIDIA hardware. Built on NVIDIA Nemotron models and trained using synthetic voice-agent data through the NeMo framework, PhoneLLM uses supervised fine-tuning and is exploring reinforcement-learning enhancements for future versions.