Both models are fine-tunes of Nvidia Nemotron architecture, with Mini using Nemotron 3 Nano and Max using Nemotron 3.5 Super VL. The open weights are hosted on HuggingFace, and the tools are available through inference partners AWS, Baseten, and Modal. The Mini model is also available on Cloudflare Workers AI through a collaboration with the Pipecat team to streamline deployment.
Tuesday, October 20, 2026
Story
PhoneLLM Debuts Open Voice AI Models With Sub-100ms Latency
AI5 days ago
Two new artificial intelligence tools for conversational audio have arrived to improve how virtual assistants process spoken input. The PhoneLLM team has launched PhoneLLM Mini, a 30B parameter model with sub-100ms time to first audio token (TTFAT) that fits on an RTX 5090 and can run over 100 concurrent agents on an Nvidia B200. The larger PhoneLLM Max is a 120B parameter model designed for complex tasks with sub-200ms TTFAT, operating at 1/10th the latency and cost of comparable models.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.