# P-JiT — Linum's open 2B text-to-image model trained for $21K

> A research text-to-image model with open weights, built on a small budget, with the full training recipe published.

- Org: Linum
- Date: 2026-11-04
- Category: model
- Source: https://www.linum.ai/field-notes/pjit-21k
- Canonical page: https://ai-tldr.dev/releases/linum-pjit-21k/

## In short

Linum released P-JiT, a 2.18B-parameter text-to-image model that works directly in pixel space. Pretraining and post-training cost $21,038 on 8,415 H100-hours. Weights and code are Apache 2.0.

## What it is

P-JiT (Pyramid-JiT) is a 2.18B-parameter text-to-image model from Linum, released with Apache 2.0 weights on Hugging Face and code on GitHub. Linum's field notes describe how the whole model was trained for $21,038. Linum calls it a research checkpoint, not a production model.

## How it works

A frozen Qwen3.5-4B text encoder feeds a diffusion transformer that works on raw pixels in 32×32 patches, with no separate image autoencoder. Pretraining saw about 610M images at 512² and about 50M at 1024² on 64 H100s, using the Muon optimizer. A variant of Finite Difference Flow Optimization was used for post-training.

## Why it matters

Linum's write-up shows that a small team can train an image generator from scratch for about $21K, with the costs, hardware and training settings all public. The known gaps are stated: the model cannot write text in images and struggles with crowds and full-body shots.

## Who it's for

ML researchers and teams training their own image models

## Try it

git clone https://github.com/Linum-AI/pjit-21k && cd pjit-21k && pip install -e . && python generate.py --no_upsample --aspect_ratio square --seeds 0 --prompt "..."

## Links

- [Field Notes](https://www.linum.ai/field-notes/pjit-21k)
- [GitHub](https://github.com/Linum-AI/pjit-21k)
- [Model Card](https://huggingface.co/Linum-AI/pjit-21k)
- [HN Discussion](https://news.ycombinator.com/item?id=50368208)