Sapient Intelligence launches HRM-Text, challenging the LLM monopoly with a brain-inspired foundation model trained on up to 1000x fewer tokens

Launched fully open-sourced, HRM-Text delivers advanced reasoning and strong language capabilities with 1/1000 of the training tokens required by leading models, enabling deeper multi-step problem solving with far less compute

SINGAPORE, May 18, 2026 /PRNewswire/ -- Sapient Intelligence, an AGI research company, announces the launch of HRM-Text, an ultra-lean 1-billion-parameter reasoning language model, to deliver competitive reasoning and general performance without the infrastructure and GPU demands of Transformer-based LLMs. HRM-Text is built on Sapient Intelligence's novel hierarchical architecture that goes beyond standard transformer design. Instead of one stack doing a single forward pass (the standard Transformer approach), HRM-Text runs two stacks in a nested recurrence in a continuous latent space, before any output is produced. That hierarchical recurrent structure is what makes it distinct from a standard LLM.

The industry continues to scale AI toward trillion-parameter systems that require vast data center infrastructure to support models trained to predict the next token in a sequence, learning statistical patterns from internet-scale datasets. Sapient Intelligence offers a fundamentally different approach by separating reasoning from language generation, mirroring the brain's natural approach to problem solving. The result is a model that produces high-quality outputs and achieves strong reasoning and general performance using 40 billion tokens of structured data, up to 1000x fewer than the 4 to 36 trillion tokens typically required. The entire model can be trained in one day on a budget of approximately $1,000, compared to other models, which cost hundreds of millions to train. 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.

Conceptually different from the next-token prediction training approach used by current LLMs, HRM-Text takes a task-completion training approach, where the model extracts and learns similarities and patterns directly from structured tasks in mathematics, logic, and general knowledge, and embeds reasoning directly into its computation. Across benchmarks, the model remains competitive with significantly larger systems while operating at a dramatically lower training budget, demonstrating that performance does not depend on massive data and compute. This efficiency is not the result of optimization alone but of fundamentally different approaches to AI that go beyond Transformer-based large language models.

A core challenge of traditional LLMs is that complex problem-solving is often expressed through long chains of generated text. By reasoning in latent space and executing multiple internal recurrent steps before generating an answer, HRM-Text adjusts the depth of its reasoning to the complexity of the task without increasing its size. This allows a compact 0.6 GiB system that fits on any modern smartphone to compete with models several times its size by reusing computation rather than scaling parameters.

Optimal for offline deployment, HRM-Text allows sensitive data to be processed locally rather than sent to the cloud, an advantage for sectors that rely on privacy, reliability, and real-time decision making. The architecture can support applications across healthcare, finance, drug discovery, and climate prediction, including nutrition and weight management support, faster molecular screening, and planning for energy, agriculture, and disaster risk management.

Sapient Intelligence is led by a team of researchers and engineers from leading AI labs, including DeepMind, DeepSeek, and xAI, alongside Neuroscience experts from the Tsinghua Laboratory of Brain and Intelligence and top institutions such as MIT, the University of Cambridge, Carnegie Mellon University, Tsinghua University, the University of Alberta, and Peking University. The model launches today on GitHub and is fully open-sourced.

"The current AGI landscape has been defined by scale, but scale alone isn't solving the real limitations of the industry," says Guan Wang, CEO of Sapient Intelligence. "With our newest HRM-Text model, we set out to show that a more capable AI does not have to come from bigger models or even greater compute demands. By building around efficiency, we believe industry progress will come from how the system thinks, not just how much it consumes. The brain has been refined over millions of years to solve complex problems with extreme efficiency, using just 20W of power and approximately 1B learned tokens. That principle is central to how we believe the next generation of AI should be built."

Model Specifications:

• HRM-Text is a 1B-parameter hierarchical reasoning language model designed to deliver strong reasoning and language performance with a much smaller training and deployment footprint.
• In an independent April 2026 verification, HRM-Text achieved 56.2% in MATH, 81.9% on ARC-Challenge, 82.2% on DROP, and 60.7% on MMLU (Benchmark explanations below).
• HRM-Text was trained on roughly 40B effective tokens, up to 1000x fewer than the 4T to 36T tokens used in many contemporary pretraining runs.
• This task-completion training approach helps optimize the model for reasoning depth rather than broad memorization, enabling strong performance with a smaller training budget.
• HRM-Text performs reasoning in a continuous latent space, allowing the model to complete internal recurrent steps before producing an answer. This reduces reliance on long visible reasoning traces and large output-token budgets.
• Each forward pass includes 8 internal recurrent steps: 2 high-level and 6 low-level updates, creating a multi-timescale reasoning process that increases computation per token without increasing model size.
• HRM-Text was pretrained in approximately 1 day on a smaller budget of just about $1,000 (Exact numbers may vary across different environments). The model uses 16 GPUs across two machines, significantly lowering the compute requirements for developing advanced reasoning models.
• At int4 quantization, HRM-Text occupies approximately 0.6 GiB, making it suitable for on-device and offline deployment within the memory budget of contemporary smartphones and edge devices.
• With the ability to deliver strong performance fully offline, HRM-Text can support privacy-sensitive use cases where data control, latency, and offline access are important.

Benchmark Explanations:

• MATH: A benchmark that tests mathematical reasoning and problem solving, often requiring multi-step logic rather than simple recall.
• ARC-C: The AI2 Reasoning Challenge- Challenge Set, designed to test science reasoning through difficult grade-school science questions that require inference and commonsense understanding.
• DROP: A reading comprehension benchmark that tests a model's ability to reason over passages, especially with numbers, counting, comparison, and discrete operations.
• MMLU: Massive Multitask Language Understanding, a broad benchmark covering many subjects, used to evaluate general knowledge and multi-domain reasoning.

Use Cases:

HRM's distinctive architecture—characterized by its deep reasoning capabilities and high learning efficiency—enables the model to excel at complex, long-horizon tasks while maintaining a remarkably small form factor. This efficiency unlocks new opportunities in high-impact sectors where conventional transformer models often falter or become cost-prohibitive to train. Building on this foundation, Sapient Intelligence develops specialized domain intelligence powered by HRM to deliver measurable performance gains across key fields, including embodied AI, quantitative finance, healthcare, scientific research, and climate prediction.

• In embodied AI, the models support vision-language-action (VLA) systems that help machines better understand and respond to real-world environments.
• In quantitative finance, they can be applied to wealth growth.
• In healthcare, the models support nutrition distribution and weight-management applications by providing more personalized insights and recommendations.
• In AI for science, the technology can accelerate drug screening by more efficiently identifying promising molecular candidates.
• In climate prediction, the models can support forecasting and decision-making across energy, agriculture, and disaster risk management.

About Sapient Intelligence

Sapient Intelligence is pursuing Artificial General Intelligence (AGI) by developing a next-generation, brain-inspired AI architecture that overcomes the structural limitations of traditional AI frameworks. By integrating reinforcement learning (RL), evolutionary algorithms, and neurodynamic principles, Sapient Intelligence develops models with advanced logical reasoning, lifelong learning, and high interpretability.

In June 2025, Sapient Intelligence introduced the Hierarchical Reasoning Model (HRM). HRM demonstrated exceptional reasoning performance across complex tasks, including advanced mathematics and Sudoku puzzles, and outperformed leading models such as DeepSeek R1 and OpenAI o3 in the ARC-AGI Challenge, despite having tens of thousands of times fewer parameters.

Building on these advancements, Sapient Intelligence is extending its architecture to high-impact domains such as healthcare, AI4S, quantitative finance, climate prediction, and embodied AI, where robust reasoning, adaptability, and efficiency are critical.

The Sapient Intelligence team consists of over 40 members, with its core technical group composed of scientists and engineers from leading AI organizations such as Google DeepMind, Xai, and DeepSeek. Sapient Intelligence continues to advance AGI across borders, further expanding its operations across its Singapore, Beijing, and Palo Alto offices. To learn more, visit: https://sapient.inc/

Media contact:
Alona Stein
ReBlonde for Sapient Intelligence
[email protected]

SOURCE Sapient Intelligence