A Singapore-based AGI research company called Sapient Intelligence has released the full weights, pretraining code, and data pipeline for HRM-Text, a language model built on what it calls a Hierarchical Reasoning Model architecture. The entire package is available under the Apache 2.0 license on Hugging Face and GitHub, which means anyone can use it, modify it, or build on it commercially.
The notable part isn’t just the open-source gesture. It’s the economics. HRM-Text, which has approximately 1 billion parameters, was pretrained on around 40 billion tokens and cost an estimated $1,000 to $1,500 on modest hardware. For context, frontier models from the major labs typically train on trillions of tokens using compute budgets that run into tens or hundreds of millions of dollars.
How it actually works
The model uses what the company describes as a hierarchical latent recurrent architecture. The high-level module handles slow, abstract reasoning, cycling through 2 “H-cycles” per pass. The low-level module handles fast, detailed processing, running 3 to 6 “L-cycles” per pass. Together, these modules allow the model to iterate on its own reasoning multiple times before producing an output, all while keeping the parameter count fixed.
Rather than generating visible chain-of-thought tokens that eat up context window and inference costs, HRM-Text does its iterative thinking internally. The reasoning happens in the model’s hidden states, not in the output text.
Benchmark performance that punches up
Independent evaluations conducted in April 2026 tell the story in numbers. HRM-Text scored 56.2% on the MATH benchmark, 82.2% on DROP, 81.9% on the ARC-Challenge, and 60.7% on MMLU. The pretraining reportedly completed in around one day.
HRM-Text builds on Sapient’s earlier work. The original HRM model, open-sourced in July 2025, had just 27 million parameters and demonstrated strong performance on complex reasoning tasks with as few as 1,000 training examples. HRM-Text represents a roughly 37x scale-up in parameters while maintaining the same architectural philosophy.
Sapient has set up a Discord channel for developers interested in exploring the architecture. The Apache 2.0 license removes the legal friction. The $1,500 training cost removes most of the financial friction.