PhoneLLM launches open voice-AI models with sub-100ms latency, tops PhoneBench at 83%
charles_irl · x · 2026-10-16
kwindla releases two PhoneLLM models optimized for conversational voice AI, both full-weights fine-tunes of NVIDIA Nemotron models with open weights on Hugging Face.
- PhoneLLM Mini (30B): based on Nemotron 3 Nano, targeting lowest latency (sub-100ms TTFAT), size, and cost; solid multi-turn conversation and tool-calling, runs 100+ concurrent agents per B200 and fits on an RTX 5090.
- PhoneLLM Max (120B): built on the new Nemotron 3.5 Super VL, a "microthinking" model for complex conversational tasks with sub-200ms TTFAT, claiming SOTA task completion at 1/10th the latency and cost of comparable models.
- The attached PhoneBench v1.1 leaderboard shows Max Alpha 2 at 83.0%, ahead of Claude Sonnet 5.5 (80.3%) and Gemini 3.6 Flash (73.9%).
- Available via four inference partners: AWS, Baseten, Cloudflare, and Modal.
Related event: PhoneLLM Open-Sources Two Voice Models with Sub-100ms Latency(2 posts)→
More from Models
- Nemotron 3.5 Super: 120B params, 12B active, Mamba-2 hybrid MoE with 1M context — ollama · 2026-10-16
- NVIDIA's Nemotron 3.5 Super VL lands on Ollama with up to 5x single-GPU agent throughput — ollama · 2026-10-16
- Dev pushes back on 'dsv4.1 flash ~= Opus 4.8' claim: not close in interactive work — mertdumenci · 2026-10-16
- Meta's Muse Wonder lands in Cursor, tops Muse family at 46.1% on CursorBench — AIatMeta · 2026-10-16
- ChatGPT outage logs users out across all mobile apps as OpenAI investigates — JeremyNguyenPhD · 2026-10-16
- Meta open-sources Muse Spark Open agent model with day-0 TPU support in SGLang — BanghuaZ · 2026-10-16