Linum published P-JiT (Pyramid-JiT) on November 4, 2026. It is a text-to-image model whose full training run cost $21,038 in GPU compute, according to the company's field notes. The weights are on Hugging Face at Linum-AI/pjit-21k, and the code is on GitHub, both under an Apache 2.0 research license.
The model has 2.18 billion parameters at sampling time, plus about 0.24 billion more in heads used only during training. It works directly in pixel space and uses a frozen Qwen3.5-4B text encoder. The architecture uses 22 DiT trunk blocks and 32x32 patches, which Linum says cuts the attention context to a quarter of its earlier Linum v2 model's.
Training took 8,139 H100-hours for pretraining and 276 H100-hours for post-training, 8,415 H100-hours in total, priced at $2.50 per hour. It ran on 64 H100 GPUs across 8 nodes. The model saw 660 million training samples, about 610 million at 512x512 resolution and about 50 million at 1024x1024, with a global batch size of 2,048. Linum says P-JiT converges in 11 times fewer samples than Linum v2. The post does not report standard benchmark scores such as FID or GenEval.
Linum is direct about the model's limits. Images containing rendered text were filtered out of training, so the model cannot produce readable text or posters. Linum also says it struggles with full-body figures, scenes with many entities, interactions between multiple people and world knowledge. Linum, run by co-CEOs Sahil Chopra and Manu Chopra, describes P-JiT as a stepping stone toward larger video models and says a bigger model, such as a 7B version, plus more post-training should address some of these weaknesses.
For researchers and small teams, the release offers a fully documented training run with open weights and code. Training a usable image model from scratch is no longer only possible for large labs, and the detailed cost breakdown is a useful reference for budgeting similar work.