Mithil Vakde trained a small transformer from scratch and scored 44% on the ARC-AGI-1 public evaluation set. The run took about 1.5 hours on one consumer graphics card. He puts the full training and inference cost at 67 cents. The same model scores 7% on the harder ARC-AGI-2 set. The code is open source. He says the score matches two recent recursive models, TRM and HRM, without using recursion. Gains came from a more modern architecture, more diverse data, and training on output tokens only. Cost fell through fewer data augmentations and a different optimizer. Tests removing the position and per-task embeddings dropped scores to around 24%. He argues the benchmark should ban synthetic data and offline pretraining.
What changed
His earlier version of the model scored 40% at higher cost.
What it unlocks
Reproducing ARC-AGI results on a single consumer GPU for under a dollar.
- 44% on ARC-AGI-1 public eval
- 7% on ARC-AGI-2
- 67 cents total compute cost
- 1.5 hours on one RTX 5090
What you need to act on it
- a single high-end consumer GPU or rented equivalent
- willingness to run the open-source code
Sources