PhoneLLM Mini and Max arrive as open-weight models for conversational voice AI

The PhoneLLM team says 30B Mini can run more than 100 concurrent agents on a B200 and fit on an RTX 5090. It describes 120B Max as a “microthinking” model for complex conversations, claiming task completion performance at one-tenth the latency and cost of similarly capable models. It says both are Nvidia Nemotron fine-tunes, available as open weights on Hugging Face and through AWS, Baseten, Cloudflare and Modal.