Sapient Intelligence releases HRM-Text, a hierarchical 1B-parameter model trained for $1,000.

Sapient Intelligence released HRM-Text, a 1B-parameter hierarchical model trained for $1,000 on 40B tokens that matches larger models on key benchmarks.

Sapient Intelligence, a Singaporean startup, has released HRM-Text, a 1-billion-parameter language model distributed under an Apache 2.0 license. The model deviates from the classic Transformer architecture: it is based on HRM (Hierarchical Reasoning Model), a two-module recurrent scheme inspired by brain function. A so-called high-level (slow) module handles abstract reasoning and planning, while a low-level (fast) module processes detailed computations. The model "thinks" in a loop within a latent space, following a logic that aggregates System 1 and System 2 modes.

The training figures are as unusual as the architecture. HRM-Text was trained on only 40 billion tokens (English only), approximately one-thousandth of the data volume used by comparable models, in 46 hours on 16 H100 GPUs, at a cost of approximately 1,000 dollars. The XL version claims 84.7% on GSM8k, 56.5% on MATH, 82.3% on DROP, 60.7% on MMLU, 81.9% on ARC-C, 63.4% on HellaSwag, 72.4% on Winogrande, and 86.2% on BoolQ.

This is a pre-alignment checkpoint: no instruction or chat fine-tuning is included; the model must be adapted by its users to serve as an assistant. The training, evaluation, and conversion code is available on GitHub, and the checkpoint on Hugging Face. Sapient accompanied the release with a video tutorial covering training and fine-tuning on a single GPU.

HRM-Text extends work initiated with HRM, a 27-million-parameter pure reasoning model that outperformed much larger models on certain benchmarks. The team claims the goal of opening up the training of high-performance models beyond large laboratories.