"An image model trained for about twenty thousand dollars" is a striking sentence, so it deserves a careful read. On November 4, 2026, Linum released the 2B text-to-image model P-JiT 21K with code and weights under Apache-2.0. This article covers what the cost covers, how quality is reported, and what to watch when reusing the number.

The cost breakdown

According to Linum's write-up the total is $21,038.

Stage H100 hours Cost
Pretraining 8,139 $20,348
Post-training 276 $690
Total 8,415 $21,038

The rate is $2.50 per H100-hour. Pretraining ran on 64 H100s over 610M samples at 512px and then 50M at 1024px (660M total), and restarts are included. The arithmetic holds: 8,415 × $2.50 = $21,037.50.

The Hugging Face model card says only "about $21K" and gives no GPU hours. The detail lives in Linum's blog post, so cite that.

What the number does not include

The post does not say what, if anything, is excluded. If you plan to build something similar, count these separately. This is my checklist, not a list of exclusions Linum has announced.

  • Data collection, filtering and captioning
  • Exploratory runs and hyperparameter search beyond restarts
  • Human time for evaluation
  • Salaries

How quality was reported

The only quantitative metric is validation loss on a fixed set of 20,480 square images. There are no GenEval-style benchmark scores; quality is shown through a qualitative gallery. So the question "how good is it for the money" cannot be answered from this post alone. To compare it with another model, generate from the same prompts yourself.

Stated limitations

  • Rendered-text images were filtered out of the training data, so it cannot write text or make posters.
  • It is a small model (2B DiT, 4B text encoder) that struggles with full-body figures, many entities, multi-person interactions and world knowledge.
  • The authors expect a larger model (for example 7B) and more post-training to help.

Why video comes up

The authors call the run a proof of concept toward their next video model: it tests whether the Pyramid-JiT architecture holds up under longer training. Attention cost (O(n²)) is the main bottleneck in video training, which is the motivation. No video results exist yet, so that part is still a hypothesis.

Practical judgment

  • As a production image model, today's limits (text, crowds, full-body) narrow the use cases.
  • As a research and learning artifact, a documented, low-budget from-scratch run with open code and weights is valuable, and you can try to reproduce it.
  • When comparing training costs, do not stop at GPU rental. Add the checklist above, because the same "$21K" with a different scope is not comparable.

If the broader workflow of producing writing, images and video with generative AI is what you want to learn, the course below is the related one. For the habit of comparing results carefully, see how to read AI benchmark scores.

Conclusion

The value of P-JiT 21K is less that it was cheap and more that its accounting is public enough to check. The cost is GPU rental, and quality is shown through validation loss and a gallery rather than benchmarks. When you quote the figure, quote its scope with it.

Sources