2-person lab releases 2B pixel-space text-to-image model trained for ~$20K

liuziwei7 · x · 2026-11-05

Linum, a 2-person lab, released a 2B Pyramid-JiT text-to-image model as open weights (Apache 2.0), trained from scratch for only $20K. Their pixel-space architecture converged 11x faster than a latent-diffusion counterpart, positioning it as a milestone toward affordable video models. Commenters read it as evidence that architecture conventions haven't converged yet.

Related event: Two-person lab trains 2B text-to-image model from scratch for $21K, open-sources it(3 posts)→

Original post →

More from Multimodal

Multimodal channel →