RPM‑3.3

RPM (Representations of Purpose in Music) is our proprietary music–text model. RPM learns from what the world actually does with music. It’s the first model trained on observed use, watching millions of short-form videos giving music context beyond sound.

Scroll
The missing axis

Not what it sounds like.
What it’s for.

Genre organised the record store. Streaming buried it. Music–text models still learn from genres and descriptions of sound. Listeners already moved on to purpose, sorting music by playlist instead of genre.

3%
of use explained by acoustic similarity
Observed use

Trained on watching. Not listening.

Where other models read captions about sound, RPM learns from what the world does with music, watching millions of short-form videos, the surface where songs break first.

Capabilities

Purpose-built. Literally.

Search your catalog by the language sync briefs and playlists are already written in. Or zero-shot tag your catalog by the labels your team actually need.

Results

Top of the charts, on purpose.

On the standard public benchmark closest to purpose RPM‑3.3 leads every model we evaluated. On generic tagging it matches the strongest permissively licensed baseline.

RPM‑3.3
MuQ-MuLan
LAION-CLAP
Responsible release

Generates nothing. Finds everything.

Supply is not music’s problem, most tracks are never found. RPM is built entirely on the demand side, not to add AI slop to the supply.

Discover music

Given a purpose in plain language, RPM retrieves the recordings that serve it.

Generate music

RPM is an encoder with no decoder, therefore cannot generate audio.

Train GenAI models

We don’t license RPM for use as a conditioning signal for generative audio models.

Genre had a century. Purpose starts now.

Commercial licensing available for record labels, publishers, platforms and all catalog owners. Tell us what your catalog needs to answer.

RPM‑3.3 · Available for license
Contact Us Read the Preprint