Linum Trains a 2B-Parameter Text-to-Image Model for $21,038 and Releases It Under Apache 2.0

Linum, a two-person research lab, released Pyramid-JiT 21K on November 4, 2026, a text-to-image model of about 2 billion parameters that its authors say cost roughly $21,000 in compute to train. The weights are under Apache 2.0.

The figures come from Linum's field notes post and the Hugging Face model card. Both are Linum's own. The post lists Sahil Chopra and Manu Chopra as co-CEOs.

The $21,038 breakdown

Linum priced H100 time at $2.50 an hour. Pretraining took 8,139 H100-hours for $20,348, and post-training took 276 hours for $690, for a total of 8,415 hours and about $21,038. Pretraining ran on 64 H100s in 8 nodes, with restarts counted.

The model saw 610 million samples at 512 pixels and about 50 million at 1024 pixels, 660 million in all. Images containing rendered text were filtered out of training.

1. Pretraining8139 H100-hours
2. Post-training276 H100-hours

Source: Linum field notes, November 4, 2026 (vendor-supplied), accessed 2026-11-05

What the model is

Item	Linum's figure
Diffusion transformer	2.18B at sampling
Text encoder	Frozen Qwen3.5-4B, downloaded separately
Patch size	32x32 pixels
Weights file	8.7 GB, fp32
Output sizes	1024x1024, 864x1216, 1152x864, 1248x832, 1280x720

It is a pixel-space model, so there is no separate image autoencoder. Linum says the 32x32 patches cut attention context by 4 times against its earlier 16x16 approach. The authors wrote that the design converged "in 11x fewer samples." That is a comparison with Linum's own earlier model.

Post-training used a variant of Finite Difference Flow Optimization with human-in-the-loop judges. The authors wrote, "To be blunt, post-training is a very young discipline within diffusion models."

What is missing

Linum reports no standard image benchmark such as GenEval or FID. The only metric in the post is validation loss on a fixed set of 20,480 square images, which is not a measure of how images compare with another model's.

The model card calls the release "an experimental research artifact, not a production model." It says "This model cannot write text or generate posters," and that it struggles with full-body shots, crowds and multi-person interactions. The card gives no VRAM requirement.

The Hugging Face page showed 16 likes and 68 downloads in the last month when The Terminal read it. The post also says Claude Opus 5.5 helped build the gallery and diagrams and wrote the model card and repository documentation.

Context for the $21K figure

The number covers pretraining and post-training compute only. It excludes failed experiments before the final run, staff time and data acquisition, and the post does not give a figure for any of them. For scale, $21,038 is about two and a half months of the $100,000 to $200,000 a year token bill one heavy agent user reported, taken at the low end.

It is also a long way from the scale of Kimi K3.1 Pro at 2.8 trillion parameters.

Linum calls the release a stepping stone to larger video models and expects a 7B version with more post-training to help. It gave no date.