Build for Your World · Meta Global AI Developer Hackathon

ToyVentor

Helping the next generation grow into robotics through play. Kids design and build toys, robots and upgrades, then play with them across digital, physical and AR worlds.

FIG 01Watch the pitch

3:22 · 16:9 · H.264

Pitch video coming soonIn the meantime, play the alpha or read how it works.

Try the PoC

FIG 02In the demo

Three things to look for

A photo becomes a toy

Point the camera at a spoon. SAM 3.1 outlines it and Kip picks it up.

A world the child describes

Describe a world and custom Glimmer writes the level, while Muse Image paints its sky and backdrop. Then it is ready to play.

A custom-trained Glimmer

Meta's open-weight model, custom trained by us and hosted on Modal. 54% faster than untuned, 98.8% of parts printable, 68 of 70 red-team prompts passed.

FIG 03Use of Meta tooling

What each Meta model does

Model	Role in ToyVentor
SAM 3.1	Turns a photo into geometry: a toy, a room or a held object becomes a game shape and a print shape.
Muse Image	Paints a world's sky and backdrop from a child's wish, using only sentences our Worker has signed.
Muse Transcribe	Speech to text: a child says a wish instead of typing it, and the rest of the app works with the words.
Muse Glimmer 30B	One of Meta's open-weight models: custom trained by us and hosted on Modal. It generates the level, and it runs Make from words: 54% faster than untuned, 98.8% of parts printable, 68 of 70 red-team prompts passed.
Muse Code coding agent	Built the app: more than 10,000 model calls across Muse Spark and Muse Wonder.

Muse Wonder/Spark, Muse Image, Muse Transcribe and SAM 3.1 run through Meta Model API. Muse Code is the coding agent that wrote the app. Muse Glimmer is an open-weight model we custom train and host on Modal, one of the inference partners supporting the hackathon.

Not Meta: Gemini text-to-speech says the names, and Mureka sings the songs.

Try it yourself

Try the PoC