← All content
DemoQuestion AnsweringOn-Prem / Edge
How we use Gradium's Phonon to chat with you about your post-trained small model use case

How we use Gradium's Phonon to chat with you about your post-trained small model use case

This week Gradium open-sourced Phonon, a family of text-to-speech models under 50M parameters, and we wanted to see what a voice agent looks like when every model in it is small.

Vineeth, our Founders Associate, handed the distil labs CLI to his coding agent, which built the training data from our documentation, trained a Qwen3.5-4B and evaluated it.

You can ask the model how the platform works, or describe your AI use case and it tells you whether a small model could handle it, with a one-line reason why. Phonon-31m then speaks that answer back.

Both models are small enough to run on a laptop, and on our held-out test set the model scores 0.89 with an LLM judge, ahead of the 0.80 of Kimi K3, the much larger teacher model it learned from.

Vineeth’s own hands-on time was under an hour; training ran in the background, and the voice demo was working the same evening.

Want a small model for your own agent? Train one for free, or read the docs this one was trained on.