Refinery

About Refinery

Refinery is Bittensor subnet 125, an incentivized research network focused on improving the algorithms behind language-model training, starting with optimizers. It connects a global pool of researchers and contributors who propose, test, and validate new optimizer algorithms under standardized conditions. Successful improvements—those that outperform established optimizers like AdamW and Muon on fixed compute and data budgets—are rewarded, and all methods, code, and results are shared openly to strengthen the open-source AI ecosystem. The project aims to accelerate advances in open large language models by making algorithmic progress a collaborative, inspectable, and incentivized process.