Linum trained a 2.18B text-to-image model for $21,038 under Apache 2.0

A line-item budget for training a frontier-ish model is still rare. The Linum lab, a two-person research group, published Pyramid-JiT on 4 November 2026. The team says it trained the 2.18-billion-parameter text-to-image diffusion transformer for $21,038. The weights are on GitHub and Hugging Face under Apache 2.0.

Linum shared the post on Hacker News as a Show HN. The thread reached 119 points on 5 November 2026, per the candidate file.

The budget, line by line

Pretraining cost $20,348, Linum says. Post-training added $690. The total of $21,038 covers 8,415 H100-hours at $2.50 per hour. Training ran on 64 H100 GPUs across 8 nodes, using bf16 precision and FSDP2.

The run used 660 million samples in total. Linum reports 610 million at 512 by 512, then 50 million at 1024 by 1024. The model uses a frozen Qwen 3.5-4B as the text encoder. The team says their 32 by 32 token reduction is four times more aggressive than Linum v2. The team also says the new model reached the same validation point with 11 times fewer samples.

The team filtered text-heavy data out of the training set on purpose. As a result the model cannot render text or make posters, Linum says. The release shows 249 sample images on the page. The subjects are people, animals, buildings, products, landscapes and stylised scenes. Linum publishes a validation loss chart. The post gives no comparative benchmark against FLUX, Stable Diffusion or any other named model.

What the release lets a buyer check

Linum calls Pyramid-JiT a stepping stone to video. That framing is Linum's. For a reader building image generation today, the stepping-stone claim is the least interesting part. The budget is the news.

Three things a team can now check that it could not check from a vendor demo. One, that a 2.18-billion-parameter diffusion transformer can be trained to publishable quality on 64 H100 GPUs. Two, that the open-weights route exists at a price a single engineering team could put through a finance review. Three, what the dataset-filtering choice costs at inference: no text, no posters.

A buyer should want a direct comparison with FLUX or Stable Diffusion. Same sample count, same resolution. Linum has not published that. Until it does, the $21,038 figure is a cost receipt, not a performance claim. The two are often confused. Linum is clear in the post that the second has not been measured.

A model under Apache 2.0 can be fine-tuned, distilled or wrapped in any commercial product. A team that wanted custom character consistency, or an in-house brand filter, now has a real base to start from. Hugging Face hosts the weights. The code is on GitHub. Linum's two co-CEOs, Sahil and Manu Chopra, run the company as an animation tools lab.

Source

Linum, "Training a Text-to-Image Model for $21K". Hacker News discussion.