Singapore Company Builds Tiny AI Model That Outsmarts OpenAI o3 and Deepseek R1

A Singapore-based startup has introduced a brain-inspired artificial intelligence model that challenges the prevailing trend of scaling up large language models.

Credit: WINSOME MARKETING

Sapient Intelligence’s Hierarchical Reasoning Model (HRM) achieves superior performance on complex reasoning tasks using only 27 million parameters and 1,000 training examples—without pre-training or chain-of-thought supervision.

The model’s architecture mimics human cognition by separating strategic planning and detailed computation into two modules. The high-level (H) module handles abstract reasoning, while the low-level (L) module performs rapid, detailed processing.

This structure enables what Sapient calls “hierarchical convergence,” allowing the model to reach stable solutions through multiple processing cycles without the computational burden of traditional recurrent networks.

On the Abstraction and Reasoning Corpus (ARC-AGI), HRM achieved 5% performance, outperforming models like OpenAI’s o3-mini-high, DeepSeek’s R1, and Claude 3.5 8K, all of which use significantly more parameters.

HRM also demonstrated near-perfect accuracy on complex Sudoku puzzles and 30x30 maze pathfinding tasks, where state-of-the-art models typically fail. It operates 100 times faster than large language models while using a fraction of the resources.

The model draws from neuroscience principles, including hierarchical processing, temporal separation of brain rhythms, and recurrent connectivity. These features allow HRM to maintain stability and depth without the training challenges of deep recurrent networks.

The release comes as the AI industry faces rising costs and diminishing returns from scaling. OpenAI’s training and inference costs are projected to reach USD 7 billion in 2024, while Google DeepMind’s models require months of training on thousands of chips.

HRM offers an alternative by restructuring computation flow rather than increasing model size. Sapient CEO Guan Wang said the model delivers genuine reasoning, not just statistical pattern matching.

The model’s interpretability also sets it apart. Unlike chain-of-thought prompting, which can produce incorrect answers with seemingly logical steps, HRM’s internal processes can be decoded and visualised.

With only 27 million parameters, HRM can run on edge devices, enabling real-time reasoning in robotics, autonomous vehicles, and Internet of Things applications.

Its efficiency and ability to learn from minimal data make it suitable for healthcare, climate forecasting, and other fields where data is limited and accuracy is critical.

The economic impact could be significant. If advanced reasoning can be achieved with smaller models, the need for massive computational infrastructure diminishes, potentially levelling the playing field for smaller companies and research groups.

HRM’s success aligns with a growing interest in brain-inspired AI. Research at institutions like Princeton and UCL is exploring similar approaches as traditional deep learning hits architectural limits.

Sapient has open-sourced HRM’s codebase on GitHub, inviting collaboration and signalling confidence in its architecture. Early applications in healthcare and robotics are already showing promise.

The company’s roadmap includes expanding HRM’s capabilities to more general-purpose tasks, aiming to shift the AI industry from scaling-dependent to efficiency-focused strategies.

---

• Sapient Intelligence’s HRM uses 27 million parameters to outperform larger models

• The model mimics human cognition with separate planning and computation modules

• HRM achieves 5% on ARC-AGI and near-perfect accuracy on complex tasks

Source: WINSOME MARKETING