Training a large language model from scratch is supposed to be expensive. Sapient Intelligence just did it for less than the cost of a MacBook Pro.
The Singapore-based startup released HRM-Text, a 1.15-billion-parameter language model trained on 16 GPUs over 1.9 days at a total cost between $1,000 and $1,500. The model is fully open-sourced on GitHub and Hugging Face, which means anyone can inspect, modify, and deploy it.
How HRM-Text works, and why it matters
Traditional Transformer-based models, the architecture behind GPT and its cousins, typically require training on trillions of tokens. HRM-Text was trained on roughly 40 billion structured tokens. That’s orders of magnitude less data, yet the model still posts competitive benchmark scores.
On the MATH benchmark, HRM-Text scored 56.2. On DROP, a reading comprehension test that requires discrete reasoning, it hit 82.2. Sapient positions these results against models like Meta’s Llama 3.2 3B and Alibaba’s Qwen 3.5 2B, both of which required substantially more resources to train.
