TechnoBaptist/phonellm-alpha-1
TechnoBaptist/phonellm-alpha-1 is a 30 billion parameter hybrid Mamba-Transformer mixture-of-experts model developed by Daily / the Pipecat team, with 3.5 billion active parameters and a 262,144 token context length. Fine-tuned from NVIDIA Nemotron 3 Nano 30B-A3B, it is specifically optimized for low-latency, multi-turn voice agent use cases in customer service, performing on par with larger models at significantly reduced cost and faster time-to-first-token. This model excels at accurate tool invocation without requiring 'thinking' tokens, making it ideal for real-time conversational AI applications.
Loading preview...
PhoneLLM Alpha 1: Optimized for Voice Agents
PhoneLLM Alpha 1, developed by Daily / the Pipecat team, is a 30 billion parameter hybrid Mamba-Transformer mixture-of-experts model, fine-tuned from NVIDIA's Nemotron 3 Nano 30B-A3B. With 3.5 billion active parameters and an extensive 262,144 token context length, it is engineered for low-latency, multi-turn voice agent applications.
Key Capabilities
- Low Latency & Cost-Efficiency: Achieves performance comparable to larger, general-purpose models (e.g., GPT 5.6 Terra) but with 94% lower cost and 1,300ms faster P95 time-to-first-token, crucial for real-time voice interactions.
- Accurate Tool Invocation: Specifically trained to accurately invoke tools in long, multi-turn conversations without needing 'thinking' tokens, addressing a common challenge in agentic workloads.
- High Concurrency: The Nemotron 3 architecture allows for efficient scaling on modern NVIDIA hardware, enabling high concurrent generations on a single GPU while maintaining excellent P95 latency.
- Benchmarked Performance: Evaluated using the custom PhoneBench v1 benchmark, which measures accuracy, speaking style, latency, and estimated per-minute runtime cost, demonstrating its suitability for production voice agents.
Good For
- Voice Agent Development: Ideal for building responsive voice agents for financial services, healthcare, retail, and hospitality customer service.
- Cost-Sensitive Applications: Offers a compelling solution for use cases where cost and latency are critical performance indicators.
- Self-Hosted Deployments: As an open-weights model under the BSD 2-Clause License, it can be run on custom infrastructure, allowing for greater control over optimization and data privacy.
- Tool-Use Intensive Agents: Excels in scenarios requiring precise and timely tool calls within conversational flows, even with thinking disabled.