Linum released P-JiT 21K, an experimental text-to-image model, with a November 4, 2026 training report that puts its compute bill at $21,038. The release gives researchers code and a downloadable checkpoint to examine a relatively inexpensive training run. That figure covers the reported run's compute, not a complete budget for building a model lab.
The team describes the checkpoint as a step toward its next video model. For developers, its immediate value is narrower: an inspectable image-generation experiment with disclosed training settings and explicit limits. MrKeyoor read the report and inference documentation but did not run the checkpoint or reproduce its sample images. This is a documentation-based account, not an image-quality test.
What the compute bill includes
Linum's cost breakdown lists $20,348 for pretraining and $690 for post-training, using a rate of $2.50 per H100-hour. The reported total is 8,415 H100-hours, with pretraining performed on 64 H100 GPUs and restarts included. Those are Linum's figures, not an independently audited invoice.
The distinction matters if you're judging whether another team could repeat the result. A compute total does not establish the full cost of data preparation, staff time or the experiments preceding this run. Nor does this release demonstrate a video model: Linum says it trained on 660 million image samples as a pressure test for its architecture.
The report describes a pixel-space design that dispenses with a separate variational autoencoder. Instead, the diffusion model learns its own compressed representation. Intermediate readouts predict lower-resolution versions of the image during training, helping supply learning signals earlier in the network. This gives researchers a specific design to investigate, beyond a cheap-training headline.
A small generator still needs a larger runtime
The repository calls it a 2B-parameter model, but that label describes the image generator. A frozen Qwen3.5-4B text encoder also processes captions. The documented runtime requires a CUDA GPU and lists 8.7 GB for generator weights in fp32 and 8 GB for the text encoder in bf16, plus activations. These storage figures are not a measured minimum-VRAM requirement.
Prompt handling introduces another dependency. By default, the supplied script sends a short prompt to Claude Opus 5.5 through OpenRouter to turn it into a dense caption, requiring an OpenRouter API key. The instructions also provide a way to supply your own caption with the --no_upsample option. That skips the external rewriting step and needs no API key. The distinction is useful for anyone assessing where their prompt goes.
Text rendering is outside this checkpoint's job
Linum says it deliberately filtered rendered-text material out of the training data because it wasn't useful for its animation goal. It warns that the model cannot write text or generate posters, and reports difficulties with full bodies, crowds and interactions between several people. Those limits should shape a trial before anyone compares a handful of attractive samples.
The code is licensed Apache 2.0, and the repository supplies sample captions, seeds and generation commands. The next useful evidence will be independent reproductions that report hardware, sampling settings and failures alongside successful outputs. Until then, the release offers a documented training experiment and a checkpoint to investigate, with image quality and practical running costs still to be checked outside Linum's own account.