Sapient's HRM-Text Model Cracks Recurrent Reasoning with $1,500 Budget

Sapient Intelligence, an AGI research company based in Singapore, has made significant advancements in recurrent reasoning by open-sourcing its HRM-Text model. The language model is built on a Hierarchical Reasoning Model architecture and was pre-trained on approximately 40 billion tokens at a cost of $1,000 to $1,500 on modest hardware.

HRM-Text's architecture allows for slow, abstract reasoning through its high-level module, which cycles through two 'H-cycles' per pass. It also has a low-level module that handles fast, detailed processing, running three to six 'L-cycles' per pass. This iterative process enables the model to refine its thinking before producing an output.

The model's performance was benchmarked in April 2026 and showed impressive results: 56.2% on the MATH benchmark, 82.2% on DROP, 81.9% on the ARC-Challenge, and 60.7% on MMLU. This achievement is notable given that it was achieved with a relatively modest budget of $1,500.