Skip to content

Recommendation systems

  • Input → outputText+HistorytoTracks
  • ModelLiquid 350M
  • Runs onCloud GPU
  • What it doesSemantic IDs · 8,165-track catalog
Open highwayClose matchRoad anthemClose matchNight driveTop matchdriving rock for road trips

Recommendwhat exists.

Personalized recommendations. Your context. Your catalog.

Or try an example:

Recommending for Rock fan, based on 30 recent tracks and 10 top artists.

Rock fan's history

Sent with every search, so the models can fit the results to this listener.

Top artists(10)

  • Foo Fighters
  • Arctic Monkeys
  • Green Day
  • Queens of the Stone Age
  • Muse
  • The Black Keys
  • Royal Blood
  • Biffy Clyro
  • The Strokes
  • Wolfmother

Recent tracks(30)

  • Everlong
  • R U Mine?
  • Know Your Enemy
  • My Hero
  • Do I Wanna Know?
  • Holiday
  • The Pretender
  • 505
  • Basket Case
  • Best of You
  • Fluorescent Adolescent
  • American Idiot
  • Learn to Fly
  • Brianstorm
  • Welcome to Paradise
  • Walk
  • Crying Lightning
  • Longview
  • Times Like These
  • Snap Out of It
  • 21 Guns
  • Monkey Wrench
  • Why'd You Only Call Me When You're High?
  • Boulevard of Broken Dreams
  • All My Life
  • I Bet You Look Good on the Dancefloor
  • Wake Me Up When September Ends
  • Rope
  • Arabella
  • She

Search, or try an example, to compare recommended tracks across the three model runtimes.

Catalog: 8,165 tracks from the public Million Playlist Dataset (RecSys Challenge 2018) — 7,143 seen in training, 1,022 held out. Same data, same trie, same catalog for every model.

How it works

The catalog becomes the vocabulary

Every track is assigned a Semantic ID, a short code whose tokens carry meaning, so similar songs share prefixes. Retrieval is then just generation: the model reads your taste and the query, and writes the IDs of the tracks that fit. This comparison uses unconstrained decoding. A trie over the catalog can optionally constrain decoding so only real, in-catalog IDs can ever be emitted: no hallucinated items, by construction. A new catalog requires adapting the model and its ID mapping.

Problem

A mood is more than a keyword. ‘Driving rock for road trips’ means something different for each listener. Recommendations need to reflect the request, recent listening, and favorite artists while finding tracks in the available catalog. Generic suggestions miss that context; familiar matches can miss the moment.

Solution

A fine-tuned Liquid model reads the request and listening history, then generates semantic IDs that map to catalog tracks. Similar tracks share parts of their IDs, giving the model a way to learn relationships within the catalog. Explore how context changes recommendations and compare Liquid with a larger model on the same music collection.

01

Read the moment.

Combine the listener’s mood, recent tracks, and favorite artists in one request.

02

Generate catalog matches.

Turn model output into track recommendations through your catalog’s semantic ID mapping.

03

Compare the compute.

See recommendations and measured latency across Liquid’s 350M model, a 1B baseline, and an optimized Liquid runtime.

Start building with Liquid

Model catalogContact usTry our API