Linum Trains 2B Text-to-Image Model for $21,038 and Releases the Weights
Summary
Linum trained Pyramid-JiT, a 2-billion-parameter text-to-image model, for $21,038 in compute: 8,415 H100-hours at $2.50 each. The two-person lab released the code and weights under Apache 2.0 and says the pixel-space design skips a separate VAE. It cannot render text.
Key Points
- Sahil Chopra and Manu Chopra, co-CEOs of the two-person lab, call the $21K run a proof-of-concept toward a larger video model.
- Pretraining runs on 64 H100s and costs $20,348, with post-training adding 276 H100-hours for $690.
- A frozen Qwen3.5-4B text encoder supplies features from layers 7, 15 and 27 to 22 DiT trunk blocks.