Linum, a two-person AI startup, has released P-JiT 21K, a 2-billion-parameter text-to-image model it says it trained from scratch for $21,038. The weights are on Hugging Face and the code is on GitHub, both under the permissive Apache 2.0 license. Co-CEOs Sahil and Manu Chopra published the full recipe on November 4.
Why it matters: training cost is usually the most guarded number in image generation. Linum gives a line-item breakdown, the hardware, sample counts and the techniques it credits for the savings. That makes this a useful reference point for teams that want to train their own image model rather than fine-tune someone else's. It is also clearly labelled as a research checkpoint, not a product.
The cost, line by line
According to Linum's write-up, all compute was billed at $2.50 per H100-hour on 64 H100s (eight nodes), with restarts included in the totals:
| Stage | H100-hours | Cost |
|---|---|---|
| Pretraining (512px, then 1024px) | 8,139 | $20,348 |
| Post-training | 276 | $690 |
| Total | 8,415 | $21,038 |
The model saw about 660 million training samples: 610 million at roughly 512px and 50 million at roughly 1024px, per the Hugging Face model card. Captions were dense, 50 to 80 words each. Linum deliberately filtered out text-heavy images such as infographics, PDFs, slides and posters.
Two things the write-up does not include: the names of the datasets, and any standard benchmark scores. Linum reports validation loss and shows a 249-image gallery, but no GenEval- or DPG-style numbers, and it does not compare cost or quality with other models.
How it got that cheap
P-JiT ("Pyramid-JiT") is a diffusion transformer that works directly on pixels instead of a compressed latent space, so it drops the separate image autoencoder most models use. Linum's main claims:
- Bigger patches. Images are cut into 32×32-pixel patches instead of the usual 16×16, which Linum says cuts the attention context by 4x. A 512px image costs as many tokens as a 256px image did in its previous model.
- Faster convergence. Linum says P-JiT converged in about 11x fewer samples than its earlier latent-diffusion model, Linum v2.
- Extra training signals. Intermediate "readout" heads predict the image at quarter and half resolution, and a small head aligns internal features to Meta's DINOv3 vision model.
- Optimizer choices. Muon on 2D weight matrices, AdamW elsewhere, and an exponential moving average that pauses during gradient spikes.
Text understanding comes from a frozen Qwen3.5-4B language model, which you download separately. Post-training used a variant of Finite Difference Flow Optimization with automated and human judges.
What it can't do
Linum is upfront about the gaps. The model cannot render text, so signs, logos and posters fail. The model card says it struggles with full-body shots, crowds, interactions between several people and world-knowledge details such as dice dots or guitar strings. The authors put this down mainly to model size and limited post-training, and suggest a 7B model would do better.
On Hacker News, where the Show HN post passed 240 points, commenters raised the same issues: hands and limbs often wrong, a smoothed "uncanny" look, and questions about cherry-picking and about where the 660 million images came from. No answer on the data source was visible in the thread when we read it.
Trying it
- Nvidia GPUs: the DiT checkpoint is 8.7GB in fp32, plus the Qwen3.5-4B encoder. Defaults are 50 sampling steps at one of five roughly 1-megapixel sizes. The bundled prompt upsampler calls an external model through OpenRouter; the model card says
--no_upsampleturns it off. - Macs: an independent MLX port, pjit-mlx, says it reproduces Linum's gallery seeds. Its author reports a 3.6GB peak in 8-bit low-memory mode and about 72 seconds per 1152×864 image on an M4 Pro in bf16. It's a brand-new, two-commit project, so treat it as experimental.
pip install git+https://github.com/felix-ab/pjit-mlx
What to watch
Linum calls this a proof of concept for its next video model and says more on its post-training stack is coming "in the coming weeks." The useful test now is independent evaluation: until someone runs P-JiT through standard benchmarks, the $21K figure tells you what it cost, not how good it is compared with other models. The data question also matters for anyone thinking of building on the Apache-licensed weights commercially.
Sources
- Linum — Training a Text-to-Image Model for $21K linum.ai
- Hugging Face — Linum-AI/pjit-21k model card huggingface.co
- Hacker News — Show HN: A decent text-to-image model trained for ~$20K (Apache 2.0 Weights) news.ycombinator.com
- AI TL;DR — P-JiT, Linum's open 2B text-to-image model ai-tldr.dev
- GitHub — felix-ab/pjit-mlx: P-JiT 21K on Apple silicon (MLX) github.com