Linum P-JiT - Text-to-image model trained for $21K

Show HN: A decent text-to-image model trained for ~$20K (Apache 2.0 Weights)

Linum, a two-person research lab, trained a 2B-parameter text-to-image model from scratch for just $21,038 using their novel Pyramid-JiT architecture. By eliminating the VAE and operating in pixel space, they achieved 32×32 token reduction—4× more aggressive than standard models—slashing attention costs. The model converges in 11× fewer samples than Linum v2 and produces surprisingly high-quality images. With Apache 2.0 weights, it's an experimental release and a stepping stone toward their ultimate goal: fast, affordable generative video tools that democratize animation.