- Reaction score
- 1,800
A tiny AI model built by a little-known startup is quietly outperforming tech giants like OpenAI and Anthropic—using just a fraction of the data and power.
A new AI architecture developed by a small Singaporean startup has made headlines after outperforming major large language models like OpenAI’s GPT-4 and Anthropic’s Claude on a notoriously difficult intelligence benchmark—using just a tiny fraction of the computational resources.
The technology, unveiled by researchers at Sapient, introduces a model called the Hierarchical Reasoning Model (HRM). According to a peer-reviewed preprint on arXiv, HRM achieved 40.3% accuracy on the Abstraction and Reasoning Corpus (ARC-AGI), a benchmark designed to test general problem-solving abilities without prior training on specific tasks. In comparison, OpenAI’s o3-mini-high scored 34.5%, Claude 3.7 came in at 21.2%, and Deepseek R1 at just 15.8%.
What’s caught the AI world off guard is the fact that HRM runs on just 27 million parameters—roughly 1,000 times fewer than mainstream models. It was trained using only 1,000 samples. No pretraining. No reinforcement learning. No fine-tuning on vast troves of internet data.
Rather than throwing more data or compute at the problem, Sapient’s team opted for something different: structure. HRM’s design mimics the brain’s ability to process information across multiple timescales.
dailygalaxy.com
A new AI architecture developed by a small Singaporean startup has made headlines after outperforming major large language models like OpenAI’s GPT-4 and Anthropic’s Claude on a notoriously difficult intelligence benchmark—using just a tiny fraction of the computational resources.
The technology, unveiled by researchers at Sapient, introduces a model called the Hierarchical Reasoning Model (HRM). According to a peer-reviewed preprint on arXiv, HRM achieved 40.3% accuracy on the Abstraction and Reasoning Corpus (ARC-AGI), a benchmark designed to test general problem-solving abilities without prior training on specific tasks. In comparison, OpenAI’s o3-mini-high scored 34.5%, Claude 3.7 came in at 21.2%, and Deepseek R1 at just 15.8%.
What’s caught the AI world off guard is the fact that HRM runs on just 27 million parameters—roughly 1,000 times fewer than mainstream models. It was trained using only 1,000 samples. No pretraining. No reinforcement learning. No fine-tuning on vast troves of internet data.
Rather than throwing more data or compute at the problem, Sapient’s team opted for something different: structure. HRM’s design mimics the brain’s ability to process information across multiple timescales.
“We’ve Created an AI That Thinks Like the Human Brain”: This Startup Beats ChatGPT With 1,000 Times Fewer Parameters
A tiny AI model built by a little-known startup is quietly outperforming tech giants like OpenAI and Anthropic—using just a fraction of the data and power.


