#Singapore #Sapient Intelligence #HRM-Text #AGI research

Sapient Intelligence Launches HRM-Text

On May 18, 2026, Sapient Intelligence, a prominent AGI research company, announced the debut of its innovative language model, HRM-Text. This model stands out in the competitive landscape of AI by employing a unique design that reduces the necessary training tokens by up to 1000 times compared to many leading systems. With just 1 billion parameters, HRM-Text is poised to challenge the conventional Large Language Model (LLM) paradigm.

HRM-Text features a novel hierarchical architecture that differentiates it from the traditional transformer models prevalent today. Instead of relying on a single stack for processing, HRM-Text utilizes a dual-stack system that executes multiple internal steps before generating any output. This method mirrors human cognitive processes and enhances the model's ability to perform complex reasoning tasks without the extensive computational demands common to typical LLMs.

The necessity for vast infrastructures and immense data centers has become a significant hurdle for the development of next-generation AI technologies. Current models often require training on trillions of tokens derived from internet-scale datasets, incurring other high costs. Conversely, HRM-Text achieves robust reasoning and impressive general performance by effectively separating reasoning capabilities from language generation, using only 40 billion structured tokens.

One of the remarkable advantages of HRM-Text is its efficiency, allowing it to complete training within a single day at a minimal cost of approximately $1,000. This contrasts sharply with mainstream models, for which training can run into hundreds of millions. Performance benchmarks reinforce the capabilities of HRM-Text: it achieved scores of 56.2% on MATH, 81.9% on ARC-Challenge, 82.2% on DROP, and 60.7% on MMLU.

What sets HRM-Text apart is its task-completion training approach, where the model learns from structured tasks and embeds logical reasoning directly into its computations. This allows it to maintain competitive performance with significantly larger models while operating on a drastically lower budget. It addresses a core challenge in traditional LLMs: long chains of generated text often hinder complex problem-solving. By reasoning in a continuous latent space with multiple internal recurrent steps, HRM-Text optimally adjusts its reasoning depth without scaling up its model size.

HRM-Text is particularly suited for offline deployment. It processes data locally, making it ideal for sectors where privacy and real-time decision-making are critical, such as healthcare, finance, climate prediction, and drug discovery. For instance, it can facilitate advanced nutritional advice, accelerate molecular screening, and assist in energy management and disaster risk planning.

Led by a dynamic team of researchers from top AI institutions, Sapient Intelligence's ambitious vision for AGI underscores the need for efficiency over mere scale. CEO Guan Wang emphasized that despite the industry's focus on expansive models, true advancement in AI should center on how these systems process information. The human brain operates on remarkably low power while mastering complex problem-solving; a principle that the team at Sapient Intelligence applies to the design of HRM-Text.

Available today as an open-source model on GitHub, HRM-Text opens up exciting new avenues for AI development. As Sapient Intelligence continues to push boundaries, the implications of this breakthrough technology across numerous domains could redefine the potential of AI, making it accessible, efficient, and highly effective in various applications.

Conclusion

The rollout of HRM-Text not only represents a leap in language model capabilities but also aligns with the global trend towards more sustainable, cost-effective AI solutions. As businesses and researchers alike explore the many applications of this innovative technology, HRM-Text could set a new standard for what AI can achieve without the need for massive computational resources.