HRM-Text: efficient training with $1k and 40B tokens

Trained on only $1,000 compute and 40 billion tokens. Uses hierarchical latent architecture inspired by the brain. Offers a full pretraining framework for accessible foundation model training.