Low-cost open text-to-image model trained for $20K
The gist
Linum, a two-person research lab, trained a 2-billion-parameter text-to-image model for $21,038 and released its weights under Apache 2.0.
The claim rests on Linum's own samples, with no benchmark against other image models.
How it unfolded
- Nov 4, 2026Linum publishes a field notes post, "Training a Text-to-Image Model for $21K," by co-CEOs Sahil and Manu Chopra.
- Nov 4, 2026VibeLeaderboard reports the model has about 2 billion parameters, is released as open weights under Apache 2.0, and cost $21,038 in compute: $20,348 for pretraining (8,139 H100-hours on 64 H100 GPUs, restarts included) and $690 for post-training (276 H100-hours), at $2.50 per H100-hour.
- Nov 4, 2026The training run uses 610 million samples at 512 pixels, then 50 million at 1024 pixels, followed by a short post-training stage using a variant of Finite Difference Flow Optimization with JEV-Image and human reviewers as judges.
Key terms
- open weights
- Numerical values in a neural network that determine how strongly inputs contribute to its calculations.
- Apache 2.0
- A permissive free-software license from the Apache Software Foundation allowing use, modification, and redistribution without royalties.
- VAE
- A neural network architecture that learns compressed representations of data using probabilistic graphical models and variational Bayesian methods.
Source: Wikipedia