Sapient Intelligence, a Singapore-based AGI research company, has released the HRM-Text model under the Apache 2.0 license, making full weights, pretraining code, and data pipeline available on Hugging Face and GitHub. With approximately 1 billion parameters trained on around 40 billion tokens, the model cost between $1,000 and $1,500 on modest hardware. In independent benchmarks conducted in April 2026, HRM-Text scored 56.2% on MATH, 82.2% on DROP, 81.9% on ARC-Challenge, and 60.7% on MMLU. The model’s hierarchical architecture uses internal reasoning cycles within its hidden states, avoiding visible chain-of-thought tokens that consume context window and inference costs. This model represents roughly a 37x scale-up in parameters compared to the original 27-million-parameter HRM model.
Source: Read the original article
