With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
Reliquary is Bittensor Subnet 81, “the RL layer of Bittensor”. It is an open protocol and reference implementation for decentralised GRPO (reinforcement learning) training of language models.
Independent miners spend their own GPU compute searching for prompts at the edge of what the model is currently learning, the “learning frontier”. The trainer admits, ranks, verifies and rewards the best groups of rollouts, and trains on them when safety gates allow.
The key incentive
Miners are not paid for producing the most rollouts. They compete to contribute verified rollout groups the trainer can actually use. Every group contains exactly 16 trajectories per prompt, and the validator independently recomputes maths rewards and executes code test cases before anything counts.
Public evidence
Rewards are validator-authoritative and evidence is published at every boundary, with a live dashboard and published training runs on a Qwen3-4B base model.
Reliquary is Bittensor Subnet 81, “the RL layer of Bittensor”. It is an open protocol and reference implementation for decentralised GRPO (reinforcement learning) training of language models.
Independent miners spend their own GPU compute searching for prompts at the edge of what the model is currently learning, the “learning frontier”. The trainer admits, ranks, verifies and rewards the best groups of rollouts, and trains on them when safety gates allow.
The key incentive
Miners are not paid for producing the most rollouts. They compete to contribute verified rollout groups the trainer can actually use. Every group contains exactly 16 trajectories per prompt, and the validator independently recomputes maths rewards and executes code test cases before anything counts.
Public evidence
Rewards are validator-authoritative and evidence is published at every boundary, with a live dashboard and published training runs on a Qwen3-4B base model.
What exists today
Reliquary is live on Subnet 81 at version 1.0, with a pinned validator container, a live dashboard and published research at reliqua.ai.
Product
The Reliquary website offers decentralised post-training on demand: customers turn their tasks into training data for their own model. Batch generation is in early access, with reinforcement learning and evaluation coming next.
What exists today
Reliquary is live on Subnet 81 at version 1.0, with a pinned validator container, a live dashboard and published research at reliqua.ai.
Product
The Reliquary website offers decentralised post-training on demand: customers turn their tasks into training data for their own model. Batch generation is in early access, with reinforcement learning and evaluation coming next.
Team
Reliquary is built by the Reliquary team (reliqua.ai). Public development on the reliquadotai/reliquary repository is led by the GitHub accounts romain13190 and 0xgrizz.
Subnet ownership
Subnet 81 now carries the Reliquary identity under its current owner. It is a different project from the one previously featured on this page.
Community
Reliquary posts on X at @reliquary_ai and has a channel in the Bittensor Discord.
Team
Reliquary is built by the Reliquary team (reliqua.ai). Public development on the reliquadotai/reliquary repository is led by the GitHub accounts romain13190 and 0xgrizz.
Subnet ownership
Subnet 81 now carries the Reliquary identity under its current owner. It is a different project from the one previously featured on this page.
Community
Reliquary posts on X at @reliquary_ai and has a channel in the Bittensor Discord.