Linum trains P-JiT text-to-image model for $21K

P-JiT is less a commercial challenger than a compelling proof that small teams can now run serious generative-model experiments on modest budgets.

• –The reported bill covers 8,415 H100-hours across pretraining and post-training—not labor, data, or broader startup costs.
• –Its pixel-space architecture removes the VAE and uses aggressive 32×32 patch compression, targeting the attention bottleneck that makes video models expensive.
• –Apache 2.0 code and weights make the result unusually reproducible for an from-scratch image-model release.
• –The model cannot reliably render text and struggles with full bodies, crowds, multi-person scenes, and world knowledge.
• –The real test is whether these efficiency gains transfer to Linum’s planned video models.