SYLIN

PhoneLLM

by Pipecat · Model · Language & code models

Fine-tune of Nemotron 3 Nano (30B, 3.5B active) for phone voice agents, open weights, no thinking, temperature 0

AccessOpen weightsNeeds~74 GBSize30B parametersLicenseBSD-2-Clause plus NVIDIA Nemotron Open Model License

Evidence

  1. Aug 30
    In a three-tool test with a rejected slot and mid-sentence party-size change, it called check availability with four, in about a second
  2. Aug 30
    Claimed to match or beat GPT-5.6 Terra and Gemini 3.6 Flash on PhoneBench, 5-10x faster to first token at a fraction of the cost
  3. Aug 30
    Shown calling create_reservation where base Nemotron 3 Nano only talks and never books the table

Cost to run