Summary
Fahd Mirza walks through a complete local installation and live evaluation of PhoneLLM, the fine-tuned voice agent model released by the Pipecat team, built on Nvidia’s NeMo-Tron 3 Nano (30B parameters, 3.5B active). The model is specifically optimized for real-time phone call pipelines — not chat interfaces — where the entire speech-to-text, LLM inference, and text-to-speech loop must complete in under 1.5 seconds to feel natural to a caller.
Mirza serves the model using vLLM on Ubuntu with a single GPU, consuming approximately 74GB of VRAM at full KV cache (with quantized options available for lower-VRAM setups). The testing sequence escalates in difficulty: first a simple reservation booking, then a multi-tool scenario where the caller changes party size mid-sentence and a previously rejected time slot is already in context. PhoneLLM correctly fires the right tool call with updated parameters in approximately one second at 640 tokens of context — the key capability voice agent builders care about most, since most general-purpose models fail at reliable mid-conversation tool switching.
The video also shows benchmark comparisons positioning PhoneLLM against GPT-5.6 Data and Gemini 3.6 Flash, where it matches or exceeds performance at a fraction of the latency and cost. Mirza connects this tutorial to a prior video covering a full Pipecat + Ollama + Kokoro local voice pipeline, positioning PhoneLLM as a drop-in replacement for the LLM component. The entire workflow uses the standard OpenAI-compatible chat API, meaning integration with existing tooling requires no client-side changes.
📺 Source: Fahd Mirza · Published August 30, 2026
🏷️ Format: Tutorial Demo
