Sapient Intelligence releases HRM-Text, a 1 billion-parameter AI, for $1,500
Sapient Intelligence just dropped HRM-Text, an AI model with one billion parameters. And the crazy part? It cost only $1,500 to build.
Instead of relying on expensive setups or tons of internet data, Sapient used its own Hierarchical Recurrent Model (HRM) architecture and trained it on 40 billion tokens.
The design splits tasks into slow-evolving strategic and fast-evolving execution layers, making the model super efficient.
HRM-Text rivals larger models
HRM-Text holds its own against bigger models like Qwen and Llama, hitting solid scores on benchmarks (MMLU: 60.7%, GSM8K: 84.5%, MATH: 56.2%) while using way less data and computing power, using 100 to 900 times fewer training tokens and 96 to 432 times less estimated compute.
Now that it's available on GitHub for download, smaller teams and startups can actually afford to build custom AI without breaking the bank.
This could make advanced AI way more accessible for everyone.