From our experiments · Experiments & artifacts

Research in practice

World models, autonomous media, and characters with a presence.

5 experiments
Playable world modelsWorld models

Pixel-Art-GWM

A compact, action-conditioned model exploring how to preserve a controllable character as its world changes.

Explore experiment →
Representation learningWorld models

JEPA in Super Mario

Exploring temporal consistency, spatial memory, and action conditioning on a trained Mario model.

Explore experiment →
Our model trainingWorld models

Talking avatars

Training our own avatar models to turn a single image into a speaking character with a consistent visual identity.

Explore experiment →
Side-by-side stills of the Joker, base MiniMax H3 and the fine-tune
Model trainingGenerative media

Fine-tuning a 33B video model

A style LoRA on MiniMax H3. 132 clips, one H100, about $55, and every wall we hit.

Read the field guide →
RSS · grounded search · Gemini newsroom · anchors & correspondents · Reactor FastH3 · seconds before air · A 24-hour tech news channel that runs itself · RSS · grounded search · Gemini newsroom · anchors & correspondents · Reactor FastH3 · seconds before air · A 24-hour tech news channel that runs itself · 
Real-time generative mediaGenerative media

R24

An autonomous newsroom that selects stories and generates anchors and correspondents seconds before broadcast.

Explore experiment →
For live demos with interactive media

Try live demos.