HRM-Text (sapient.inc)

HRM-Text has been unveiled as a groundbreaking AI model that combines data-efficient training and remarkable performance in reasoning tasks. Trained on approximately 40 billion tokens, it utilizes up to 1,000 times less data compared to conventional models that sample between 4 to 36 trillion tokens. Despite its compact architecture of 1.15 billion parameters, HRM-Text competes effectively with larger models on various reasoning-heavy benchmarks, demonstrating a footprint of just 0.6 GiB when quantized to int4, allowing it to function independently without cloud resources. The significance of HRM-Text lies in its ability to deliver competitive results across several challenging benchmarks, achieving scores of 56.2% on MATH, 81.9% on ARC-Challenge, 82.2% on DROP, and 60.7% on MMLU. These benchmarks evaluate complex logical reasoning, scientific knowledge, reading comprehension, and general multi-domain understanding. The model serves various high-impact real-world applications, marking a crucial step in making advanced reasoning capabilities more accessible and efficient, particularly in resource-constrained environments.