Noesion Radarnoesion.ai
Research2026-11-04

Linum trained a 2B-parameter text-to-image model called P-JiT for $21K. It used 660M samples during training. The model is released under Apache 2.0. Code and weights are available on GitHub and Hugging Face. P-JiT builds on a pixel-space architecture that cuts attention context windows by 4×. It converged in 11× fewer samples than earlier versions. Linum published 249 sample images across categories like people, animals and landscapes. The release is a step toward larger video models.

What changed

P-JiT previously converged in 11× more samples with a 4× larger attention context window.

What it unlocks

Training a 2B-parameter text-to-image model at far lower cost than prior methods.

  • 2B parameters
  • $21K compute cost
  • 660M training samples
  • 249 sample images

Sources

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.