Inventing

You edit one frame. The model spreads that look across the entire video — and the face stays the same.

Two output modes — what the person actually said and what they meant. Runs even without a graphics card.

Top-tier performance for about US$0.03 per million tokens — and the model is published for download.

A screenshot becomes something like a Figma file again. Under 4 GB.

2.8 trillion parameters, native vision, 1.56 TB file. The weights came out on the promised date.

16 B parameters, 2.8 B active, and something rarer: the checkpoints and training recipes published.

You pinpoint what you want to remove with the brush. What comes off along with it is what separates a good remover from a bad one.

Several video models in a single subscription, with two ready-made flows — ad and film.

It's not the smartest in its class. It's the one that understands audio, image, video, and text all at once.