Realtime voice, video, and AI at global scale
The Hook
“Hey PhoneBot. Hey John, what's on your mind? What happens when you fine tune an open weights model specifically for voice agents? Oh, great question. Let me show you.”
Product Reveal
“This is Pipecat's new model, PhoneLLM, and across the things that actually matter for production voice agents, that's how reliably it can call tools, how fast it responds, and how much it costs to run, well, it outperforms most of the models used for production voice agents today. And it does so at voice-friendly latency and at a fraction of the cost. Okay, okay, hold up. This is a pretty big claim. So, how did we get here?”
Feature Teaser
“(No spoken dialogue — background synth music plays)”
Problem Agitation
“Pipecat is the most widely used framework for building voice agents... Over the last few years, as the models have grown in size, well, they've gotten slower and slower... For voice agents, where latency between turns is everything, yeah, that's a bit of a problem. Because of this, a lot of enterprises are still using older, faster models like GPT-4.1 for voice agents... Even the newest state-of-the-art small models today... they want to think a lot before they produce answer tokens. So this is a painful gap.”
Feature Teaser
“Voice needs accuracy, but with the reasoning, well, turned off. Over the last two years, we've been building voice infrastructure with Pipecat... We've been building out and training pipelines and tooling that let enterprises take open-weight models, adapt them to their own use cases and data, and run these customized models in production.”
Call to Action
“PhoneLLM is the first in a series of models that we're releasing publicly... PhoneLLM is available on Hugging Face, so you can give it a try yourself.”
Synthesizable in Remotion · key components of the ai-voice-agent-benchmark-launch template.
Morphing glowing blue/purple gradient fluid sphere with center audio level bars pulsing to speech frequency.
Render SVG filter feTurbulence/feDisplacementMap animated via frame count. Overlay vertical rounded bars scaling along Y-axis mapped to interpolate(audioAmplitude, [0, 1], [0.2, 1.0]).
Connected numbered nodes across a baseline that sequentially highlight with glowing purple rings.
Layout row flexbox with numbered circle indicators and SVG connecting lines. Use frame sequence to trigger spring scale and border color transition from #333333 to #7B61FF.
Data table rows where horizontal score bars animate width from 0 to percentage while highlighting top model.
Map dataset to rows; animate width of bar container using interpolate(frame, [0, 20], [0, scorePercentage], { extrapolateRight: 'clamp' }) with a purple spotlight background on the hero entry.Purpose. Establish credibility by talking to the agent immediately.
Execution. Split screen real-time voice orb animation.
Purpose. Back the premise with empirical comparative data.
Execution. Leaderboard table highlighting latency, cost, and tool reliability.
Purpose. Deconstruct why contemporary reasoning LLMs fail voice interactions.
Execution. Thinking token bubble animations and expanding turn latency scrubber.
Purpose. Present the training methodology and direct developers to the open weights.
Execution. Six-stage workflow timeline transitioning into Hugging Face/GitHub download cards.