Linum, which describes itself as “a 2-person research lab training our own generative video models entirely from scratch,” released a 2-billion-parameter text-to-image model on November 4, 2026 and says it cost about $21,038 to train. The company’s post, by co-CEOs Sahil Chopra and Manu Chopra, presents it as a research release and a stepping stone toward larger video models. The post’s Hacker News submission had 235 points. All figures below are the company’s.

The model

  • The model, Pyramid-JiT (P-JiT), is a pixel-space diffusion transformer of about 2.18 billion parameters at sampling and about 2.42 billion during training, which includes training-only heads. It uses a frozen Qwen3.5-4B text encoder.
  • The weights are on Hugging Face and the code on GitHub, both under the name Linum-AI/pjit-21k, under the Apache 2.0 license.

Training and cost

  • Total cost: about $21,038, from 8,415 H100-hours at $2.50 per hour, on 64 H100 GPUs (eight nodes) in bf16 with FSDP2.
  • Pretraining: $20,348 (8,139 H100-hours) over 610 million samples at 512² and 50 million at 1024².
  • Post-training: $690 (276 H100-hours) with a variant of Finite Difference Flow Optimization and human-in-the-loop judges.
  • Data: images at several aspect ratios, with images containing rendered text, such as infographics, PDFs and slides, filtered out.

Claims and results

Linum says P-JiT “converges in 11× fewer samples than Linum v2” and uses 4× more aggressive token reduction (32×32 patches versus 16×16), which it says cuts the attention context window by 4×. It writes that it was “surprised by how good the results looked given the size of the model,” supported by a 249-image gallery and qualitative description. The post reports no standard benchmark scores, such as FID or human-preference rates, and no head-to-head comparisons with other models. It shows validation loss on a fixed set of 20,480 square images, with a spike to 0.157 at 80 million samples when the noise schedule widened.

Limits the company states

  • The model cannot render text or generate posters, because rendered-text data was excluded.
  • It struggles with full-body figures, scenes with many entities, multi-person interactions and world knowledge; examples include deformed limbs and objects that merge together.
  • Linum says training a larger model, for example 7 billion parameters, and further post-training “would likely mitigate these issues,” and calls post-training for diffusion models “a very young discipline.”

Sources and notes

This post is based on Linum’s post (November 4, 2026) and Hacker News counts. Lens did not download the weights, run the model or verify the cost, and did not contact Linum. The cost is the company’s accounting of GPU hours; Lens did not see other expenses itemized.

Image: two glass prisms splitting light into a colored spectrum, by Sobolishko via Wikimedia Commons, licensed under CC BY 4.0; resized. Stand-in photo; it is not related to Linum’s model and does not show the events described.