Reliquary is an open AI lab focused on distributed post-training through Bittensor Subnet 81. Its network brings together independent compute operators to generate training data and reinforcement-learning rollouts, while validators verify the work. Selected, healthy batches can contribute to future model checkpoints. Reliquary publishes protocol information, network activity, open weights, model cards, and research so visitors can inspect its work.

Its products and tools center on training data, task environments, reinforcement learning, and model evaluation. Users can explore SFT task formats and a catalog of code, math, and logic environments, each with defined answer formats and verification methods. The platform also presents configuration workflows for custom workloads and comparisons of released model checkpoints. Reliquary-4B is a released model featured in its published evaluation results.

The workspace is in private preview, with access requiring approval; custom workloads are being developed with pilot teams. On-demand reinforcement learning and model evaluation are described as coming soon, and GPU rental is not publicly available.