Sapient Intelligence, a Singapore-based AGI startup, has unveiled something truly disruptive in the AI world: the Hierarchical Reasoning Model (HRM).
A compact, efficient, and brain-inspired architecture that delivers breakthrough performance on complex reasoning tasks with only 27 million parameters and 1,000 training samples.
Forget the massive scale of GPT-style models. HRM proves that smarter architecture can outperform sheer size — and it’s now open source.
Inspired by Human Reasoning
Unlike traditional AI models that mimic reasoning through Chain-of-Thought (CoT) prompting — a brittle, error-prone method requiring vast training data — HRM is designed from the ground up to think like a brain.
The model is based on a concept known as Cross-Frequency Coupling, where different layers of cognition operate at different speeds:
- High-level module = 🧠 Slow, strategic thinking (meta-reasoning)
- Low-level module = ⚡ Fast, detailed execution (fine-grained processing)
Together, they form a looped, recurrent system capable of making a plan, executing it, receiving feedback, and revising — all within a single forward pass.
This mimics the dual-speed thought process of System 1 vs. System 2 from cognitive science, offering flexible, real-time adaptation to task complexity.
HRM Benchmark Results: Outclasses Larger Models
Here’s how HRM stacks up — using just 1,000 training examples and no pretraining:
| Task | Claude 3.7 8K | o3-mini-high | DeepSeek R1 | HRM |
|---|---|---|---|---|
| ARC-AGI-1 | 21.2% | 34.5% | — | 40.3% |
| ARC-AGI-2 | 3.0% | 1.3% | 1.0% | 5.0% |
| Sudoku-Extreme (9×9) | 0% | 0% | 0% | 55.0% |
| Maze-Hard (30×30) | 0% | 0% | 0% | 74.5% |
HRM achieves near-perfect accuracy on tasks where other state-of-the-art models fail entirely. Its success stems not from scaling, but from smarter computation.
How HRM Works
Key Features of HRM AI Agent:
- Hierarchical Structure: Splits tasks between slow, abstract planning and fast, concrete execution.
- Multi-Timescale Processing: Mimics biological cognition by operating on different rhythmic frequencies (just like your brain’s theta and gamma waves).
- Recurrent Reasoning: Iterative updates let it internally reflect and refine — avoiding single-point failure that plagues step-by-step logic in CoT.
- No Pretraining Needed: HRM learned entirely from 1,000 input-output examples — no massive corpora or multi-GPU setups.
Real-World Applications
HRM’s compact size and remarkable data efficiency unlock high-impact use cases where compute and data are limited:
- Healthcare: Rare-disease diagnosis where signals are sparse and reasoning is critical.
- Climate Forecasting: Achieved 97% accuracy in S2S forecasting models — a huge leap in actionable climate data.
- Robotics: Lightweight enough to run onboard as a “decision brain” in real-time robotic systems.
Why It Matters
“Chain-of-thought is clever — but it’s a workaround. HRM is the real thing. It reasons like a human brain, not just predicts the next likely token,”
— Guan Wang, Founder & CEO of Sapient Intelligence
HRM is a fundamentally new approach to AGI. Instead of stacking more layers and feeding more data, it takes inspiration from nature — using evolution-tested design principles to produce real intelligence, not just language mimicry.
With its open-source release, Sapient is inviting the world to explore what happens when AI stops pretending to reason and actually starts doing it.
Get Involved
🔗 Download the code & run HRM yourself:
https://github.com/sapientinc/HRM
For developers, researchers, and AGI enthusiasts — this is your chance to build with a system that might just be the future of AI reasoning.
About Sapient Intelligence
Based in Singapore, with labs in San Francisco and Beijing, Sapient Intelligence is building the future of artificial general intelligence by fusing neuroscience, evolutionary algorithms, and reinforcement learning.
Their team includes alumni from Google DeepMind, Anthropic, xAI, and top universities like UC Berkeley, Tsinghua, and Cambridge.
