# Sapient Intelligence — Full Content > Sapient Intelligence builds efficient, brain-inspired AI architectures for deeper reasoning, latent-space intelligence, and agentic systems. Its flagship technology is the Hierarchical Reasoning Model (HRM), a latent-space hierarchical recurrent architecture that separates reasoning from language for deeper, more efficient computation. Tagline: "Leaner Architectures, Deeper Intelligence." This file contains the full text content of the Sapient Intelligence website, organized by page, to help language models cite the company and its work accurately. For a shorter index, see https://sapient.inc/llms.txt. Key facts for accurate citation: - **HRM (Hierarchical Reasoning Model)**: A latent-space hierarchical recurrent architecture with dual-timescale processing — a high-level "slower controller" for abstract, deliberate reasoning and a low-level "faster processor" for detailed computation. - **HRM-Symbolic**: A reasoning model for symbolic problem-solving. Open-sourced July 2025. ~0.027B parameters, no pretraining, no chain-of-thought, ~100x faster reasoning, edge-deployable, with 500–1000 effective layers. - **HRM-Text**: A text-generation model based on the HRM architecture. Open-sourced May 2026. 1.15B parameters, trained on ~40B tokens (up to 1000× less data than comparable models), 0.6 GiB footprint at int4 quantization. Benchmarks: 56.2% MATH, 81.9% ARC-Challenge, 82.2% DROP, 60.7% MMLU. - **Approach**: Unlike LLMs that operate reasoning within language space, HRM performs reasoning in a deeper, abstract latent space unconstrained by vocabulary, then encodes results into natural language when needed. - **Company**: Founded 2024-08-05 by Guan Wang (Founder, CEO) and William Chen (Co-founder, Head of Strategy & Management). Headquartered in Singapore, with teams in Palo Alto and Beijing. Contact: info@sapient.inc. --- ## Home — https://sapient.inc/en **Leaner Architectures, Deeper Intelligence** Building deeper intelligence through leaner architectures. ### HRM — A Latent Space Hierarchical Recurrent Architecture The HRM architecture is composed of two coupled components operating at different timescales: - **High-level ("Slower Controller")**: Responsible for abstract, deliberate reasoning. - **Low-level ("Faster Processor")**: Responsible for detailed computations. Information flows from input through a lower-level representation and a meta-representation to output. ### Models - **HRM-Symbolic**: A reasoning model designed for deep, efficient problem-solving through hierarchical modules, dual timescale computation, and latent space reasoning. - **HRM-Text**: A text generation model based on the HRM architecture, strengthened by task completion and latent space reasoning. ### About Us Sapient Intelligence is building leaner architectures for deeper intelligence. Through brain-inspired latent-space models and agentic systems, we develop efficient AI with stronger reasoning, self-evolving capabilities, greater adaptability, and enhanced interpretability. --- ## Architecture — https://sapient.inc/en/architecture **The HRM Architecture** Our brain-inspired architecture HRM powers deep, efficient reasoning through structured computation in latent space. ### What Is Latent Space? Latent space is an internal representational space where abstract structure, relationships, and plans can form before they are expressed as words. It is the internal workspace for thought — richer and more flexible than language. ### Understanding the Brain Understanding the brain is the first step toward human-like AI. The brain consists of specialized regions, each contributing to distinct functions. - **Language Processing**: Language is processed in brain regions specialized for understanding and producing words. - **Reasoning Processing**: Reasoning takes place in brain regions responsible for analyzing logic, patterns, and relationships. ### The Limits of LLMs While language and reasoning should each operate in their own regions, LLMs operate both in the language space, making reasoning shallow and resource-intensive. This confines thought to linguistic processing, even though human thinking and reasoning operate at a much deeper level beyond language space. Conventional LLMs often reason through language space. Even abstract problems must be processed step by step through token representations. This can make reasoning longer, more fragile, and more resource-intensive. ### Why It Matters Many problems require abstraction, planning, and internal structure that do not naturally fit into a token-by-token language process. Reasoning in latent space enables deeper thinking, more efficient computation, and greater flexibility. ### The HRM Approach HRM addresses this limitation by separating reasoning from language. Inspired by the brain, it enables thinking to take place in a deeper, more abstract latent space, unconstrained by vocabulary or surface-level linguistic representations. When needed, the resulting thoughts can be encoded into natural language. This enables deeper, more effective, and more efficient reasoning than today's LLMs. Our brain-inspired Hierarchical Recurrent Model (HRM) processes information through a structured hierarchy of recurrent state updates. By decoupling core reasoning from the language interface, HRM scales deep computational capabilities independently of prompt sequence length or token footprint. --- ## HRM-Symbolic — https://sapient.inc/en/hrm-symbolic Open-sourced in July 2025, HRM-Symbolic is a reasoning model built for deep, efficient problem-solving through hierarchical latent-space reasoning. ### Key Advantages - **Greater Reasoning Depth (Multi-Scale, Multi-Step Reasoning)**: 500–1000 effective layers solve complex problems through depth, not scale. - **More Perceptive (Numerical & Pattern Sensitivity)**: Strong understanding of numbers and patterns for time-series and structured data. - **Smarter with Less Data (Small-Sample Learning)**: Learns effectively from thousands of samples, not millions. - **Smaller. Faster. Stronger. (Ultra Light, Superior Performance)**: 0.027B parameters. No pretraining. No CoT. 100x faster reasoning. SOTA reasoning performance. Edge deployable. - **More Efficient (Adaptive Computation)**: ACT dynamically optimizes inference, reducing cost without sacrificing performance. ### HRM Reasoning in Action Sudoku offers a simple, intuitive way to make reasoning visible. Here, it serves as a compact example of reasoning that HRM applies in high-impact domains. The website includes an interactive Sudoku solver that runs the HRM-Symbolic model in the browser. ### Application Domains Our architecture powers advanced reasoning across complex, high-impact real-world domains. ### Reference The Hierarchical Reasoning Model paper is available on the site. --- ## HRM-Text — https://sapient.inc/en/hrm-text Open-sourced in May 2026, HRM-Text is a 1B text generation model based on the HRM architecture, strengthened by task completion and latent space reasoning. ### Key Traits - **Data-Efficient Training**: Trained on ~40B tokens, using up to 1000× less data than the 4–36T tokens used by the models we benchmark against. - **Compact Yet Powerful**: Built with 1.15B parameters while remaining competitive with models several times its size on reasoning-heavy benchmarks. - **Native Edge Reasoning**: Runs locally with a 0.6 GiB footprint at int4 quantization, enabling advanced reasoning without cloud dependency. ### Benchmarks HRM-Text is a proof-of-concept model with no post-training. The numbers reflect architecture performance alone. Despite its compact size, HRM-Text delivers competitive results across reasoning benchmarks, including 56.2% on MATH, 81.9% on ARC-Challenge, 82.2% on DROP, and 60.7% on MMLU. Benchmark breakdown vs GPT-3.5: - **MATH**: +64.8% vs GPT-3.5. Tests mathematical reasoning and problem solving, often requiring multi-step logic rather than simple recall. - **DROP**: +28.2% vs GPT-3.5. A reading comprehension benchmark testing reasoning over passages, especially with numbers, counting, comparison, and discrete operations. - **ARC-C**: -3.8% vs GPT-3.5. The AI2 Reasoning Challenge (Challenge Set), testing science reasoning through difficult grade-school science questions requiring inference and commonsense understanding. - **MMLU**: -13.3% vs GPT-3.5. Massive Multitask Language Understanding, a broad benchmark covering many subjects, used to evaluate general knowledge and multi-domain reasoning. --- ## About — https://sapient.inc/en/about - **Vision**: Inventing the lean general intelligence: AGI beyond brute-force. - **Mission**: Building deeper intelligence through leaner architectures. ### Core Values - **First Principles Thinking**: Find the fundamental truth. - **Optimize**: Never settle. - **Open AGI**: Everyone builds. ### Our Team We bring together top-tier scientists and engineers with experience across leading AI organizations, including Google DeepMind, DeepSeek, and xAI, alongside research experts from the Tsinghua Laboratory of Brain and Intelligence and top institutions such as the Massachusetts Institute of Technology, the University of Cambridge, Carnegie Mellon University, the University of Alberta, Tsinghua University, and Peking University. Advancing AGI as a global effort, Sapient operates across Singapore, Palo Alto, and Beijing. - **Guan Wang — Founder, Chief Executive Officer**: An RL expert with experience across leading research labs and AI organizations, including Tsinghua Brain and Intelligence Lab, Shanghai AI Lab, and Pony.ai. He led OpenOrca, authored OpenChat, and has a strong track record in open-source AI development. He holds a B.S. in Computer Science from Tsinghua University and was named Forbes 30 Under 30 and Hurun Under30s. - **William Chen — Co-founder, Head of Strategy & Management**: A multidisciplinary innovator with experience across robotics, embedded systems, and commercialization. He served as R&D Engineer at DJI and Hesai Technologies, and led venture development at Tsinghua Innovation Center. He graduated from Tsinghua University with the Zhaoping-Gaorong Scholarship. ### Our Progress (milestones) The Origin · Seed Round · Seed+ Acceleration · HRM Paper · HRM Open Source · Youngest-Ever Speaker at Fortune Brainstorm AI · AIMED 25 Award for Healthcare AI Impact · HRM-Text Open Source. ### Our Locations We operate where innovation happens. Across Singapore, Palo Alto, and Beijing, Sapient Intelligence connects leading global innovation hubs to build the next generation of AI. --- ## Careers — https://sapient.inc/en/careers **Shape the future of intelligence with us.** Why join us: Remote-First (work from anywhere; results over desk hours), Learning Budget (reimbursed courses, books, and conference tickets), Premium Gear (latest MacBook, monitor, and accessories), Flexible PTO (unlimited vacation with a mandatory minimum of 25 days off per year). For questions about open roles, email careers via the site. See open roles at https://sapient.inc/en/careers. --- ## News — https://sapient.inc/en/news Explore the latest news, research, and insights from Sapient Intelligence, including announcements, research blogs, and press coverage. For press inquiries, reach out to the PR team. Index: https://sapient.inc/en/news. --- ## Contact — https://sapient.inc/en/contact-us Get in touch at info@sapient.inc. Contact form: https://sapient.inc/en/contact-us. --- ## Elsewhere - GitHub: https://github.com/sapientinc — Open-source model code and repositories. - Hugging Face: https://huggingface.co/sapientinc — Released model weights. - LinkedIn: https://www.linkedin.com/company/sapientinc/ - X / Twitter: https://x.com/Sapient_Int - YouTube: https://www.youtube.com/channel/UCd1pXtPtluwfKCRZCgKL63w - Discord: https://discord.com/invite/sapient