The chart follows the alpha price in τ. Each period starts at its opening price and steps through the mechanisms below, in a fixed order, to its closing price: a step that rises pushed the price up, one that falls pushed it down, and its height is how far it moved the price. A thin line carries each close into the next period’s open; it is dashed where the two differ. The faint candle behind each period is the market’s own open, high, low and close, for context.

With many periods on screen (90 days, or hours), the steps would be too thin to read, so the chart draws the price as a line instead and colours each period by the mechanism that moved it most. Hover a period for the full step-by-step breakdown.

Protocol buy: The chain spending excess emission TAO to buy alpha from the pool. A permanent protocol buyback; it pushes the price up.
Root sell: Root-dividend claims sold into the pool; they push the price down.
User buy / sell: Staking into and unstaking from this subnet.
Cross-subnet: Swaps from other subnets routed through this pool.
Residual: What the decomposition cannot attribute to a mechanism. It is always shown, never folded into another part.

Not here, because they do not move the spot price: the per-block TAO injection into the pool, and burns.

All figures are log changes, not simple percentage changes, so the parts add up exactly to the total; on big moves the two differ (a 10% log fall is a 9.5% price fall).

Today so far

2026-11-05 UTC, as of ≈01:40 UTC (block 9,436,594)

Price, today so far: 0.004808 τ
Volume, today so far: 0.11 τ
Market cap incl. burned, today so far: 11.99K τ
TAO in pool, today so far: 15.91K τ
Emission share, today so far: —not available yet

The day isn’t over yet, so these figures keep changing. They are in no daily total, average or rank.

Subnet Configuration

Pool & Mechanism

AMM Pool

Balancer

Mechanism

dTAO

Access

Registration

Open

Neurons

256/256Competitive

Validators

6/6458 open

Economics

Owner Cut

18.00%

Weights

Commit-Reveal

Enabled (v3, 1 epochs)

Liquid Alpha

Disabled

Weights Version

Weights Range

1–65535

Timing

Tempo

360~1.2h

Immunity Period

10,799~1.5d

Activity Cutoff

5,000~16.7h

Market & Economics

Market

Market Cap

6,686.6999 τ

FDV

27.90K τ

TVL

31.83K τ

Lifetime Volume

964.88K τ

Holders

1,574

Owner Locked

203,423.7123 ᚠ(14.63% of circulating)Perpetual

14d price

Revenue / Recycle

890.23×

last day · Net positive

Emissions

<0.01%of network emissions

TAO emitted to this subnet

Sent into its pool each day: its share of network emissions

0.0041 τ/day

0.0041 τ per day

Daily distribution

Alpha paid out each day, valued at the pool price

7,200.0000 ᚠ/day

≈ 34.6205 τ·34.6205 τ per day

Miners44.21%

3,183.0649 ᚠ ≈ 15.3055 τ/day

Validators37.79%

2,720.9681 ᚠ ≈ 13.0835 τ/day

Owner18.00%

1,295.9670 ᚠ ≈ 6.2315 τ/day

Stake 1,000 τ here

170.9785 ᚠ/day≈ 0.8221 τ·0.8221 τ per day

From the validator share, before commission

84.44%staker APY

· 30d avg 84.44% · Stable

90d

Pool emits 1.0000 ᚠ/block at 0.004808 τ/ᚠ. Yield varies with stake weight, pool price, and emission rate.

For Miners

44.21% of emission

Slots

256/256Competitive

Burn Cost

0.0005 τ

Est. per Miner

5.4880 ᚠ/epoch≈ 0.0264 τ

Pool / Epoch

159.1532 ᚠ≈ 0.7653 τ

For Validators & Stakers

37.79% of emission

Validator Slots

6/6458 open

Est. per Validator

22.6747 ᚠ/epoch≈ 0.1090 τ

Registration Activity

Neuron registration trends

This Interval

24h: 17 registrations (17 burn · 0 PoW)

Target / Interval

Slots Available

Competitive

256/256

Neurons

256

Development Activity

reliquadotai/reliquaryPython

476 contributors732 commits/30dLast commit 0d ago

MIT

bittensordecentralized-aidistributed-traininggrpollm-trainingpytorchqwenreinforcement-learningsubnetverifiable-compute

Top Contributors

What Reliquary does

Language model training requires generating high-quality training examples at the policy's learning frontier — prompts where different completions produce different rewards. Traditional approaches use reactive filtering: generate many rollouts, measure variance in outcomes, discard low-variance groups. This rejects compute-expensive data post-hoc. As models improve and frontier prompts become rarer, the rejection rate rises, wasting GPU resources.

A decentralized market transforms this compute loss into competitive incentive. Independent miners predict which prompts sit at the frontier, generate rollouts at their own cost, and submit candidates to a validator. The validator verifies that submissions match announced model weights and reward claims, selects frontier groups via cryptographic proof, and pays miners based on verified contribution. Miners who pick well earn emission; miners who pick poorly absorb their own compute cost.

Reliquary implements this on Bittensor subnet 81.

Miners run the current checkpoint, generate exactly 16 completions per prompt using a deterministic forced-sampling stream, construct GRAIL sketches binding each completion to hidden-state activations, and submit signed batches. The validator independently recomputes GRAIL sketches and rewards, admits groups via continuous FIFO proof, maintains one-shot prompt cooldown by canonical-content digest, and filters for in-zone reward variance (σ ≥ 0.24). Selected groups are accumulated, trained via GRPO after quarantine checks, and published as new checkpoints to Hugging Face Hub. This ex-ante prediction market structure maintains frontier diversity while eliminating the post-hoc discard tax.

Auto-generated from the GitHub repository and may not reflect the latest changes.

About

active

Reliquary is a decentralized GRPO training system on Bittensor subnet 81 where independent miners compete to contribute verified rollout groups from frontier prompts. The protocol uses GRAIL sketches for cryptographic proof of generation, validator-authoritative rewards, and a market-driven approach to dynamic prompt sampling, aiming to match DAPO's 50% training-step efficiency with a competitive ex-ante prediction model rather than post-hoc filtering.

How it works

Reliquary operates a decentralized training loop: miners generate rollouts on their own GPUs using the current checkpoint, construct GRAIL sketches binding completions to model activations, and submit signed batches to the validator. The validator admits groups via continuous proof (FIFO), verifies GRAIL sketches and rewards, maintains one-shot prompt cooldown via canonical-content digest, and filters for in-zone reward variance. Selected groups are bundled into training payloads, accumulated across a journal of 16-slot windows, and published as new checkpoints to HF Hub.

Miners poll /state for checkpoint changes and cooldown updates, receive provisional acceptance on submit, and earn EMA-smoothed fixed-slot shares once groups are proven and selected. The validator publishes training quarantine metadata for suspected poison and maintains auditable R2 archives of every window's submissions, verdicts, and metadata.

Incentive mechanism

The zone filter (σ ≥ 0.24 for standard, σ ≥ 0.22 during bootstrap) selects groups with meaningful reward variance. Groups are ranked FIFO by upload arrival, payload metrics are telemetry-only. Each selected group earns one uniform fixed-slot share split across the environment count, picks target and B_BATCH (16).

The EMA smooths individual-window shares with α ≈ 0.027, giving ~25-window half-life. Per-epoch emissions are proportional to EMA score divided by the sum of all miner EMA scores. Miners are rejected for wrong checkpoint, bad termination, malformed proofs, GRAIL failures, prompt cooldown violations, out-of-zone σ, and fabricated completions.

The sections above were auto-generated from the GitHub repository and may not reflect the latest changes. View repository (github.com, opens in a new tab, external site)

On-Chain Identitysigned on-chain by subnet owner

ᚠ

Reliquary

Fehu · Runic letter · means “wealth”

The RL layer of Bittensor

Owner: 5GnpQ…kAQT
Hotkey: 5H47s…wdDa
Owner Cut: 18.00%
Contact: RomAI

Registered: 1y 7mo ago(Mar 25, 2025)
Block: 5,203,057
Slot history: Every subnet in SN81
Neurons: 256/256
Tempo: 360 blocks

Website

URL: https://www.reliqua.ai/ (reliqua.ai, opens in a new tab, external site)
Status: Up200
TLS: Valid
Title: Reliquary: reinforcement learning on independent GPUs
Checked: 2026-11-04 04:02 UTC

Domain registered after this subnet started · Apr 25, 2026

Changes (3)

2026-11-03 04:04 UTCUp200Reliquary: reinforcement learning on independent GPUs
2026-10-21 04:03 UTCUp200Reliquary — Decentralised post-training on demand
2026-10-13 08:00 UTCUp200Reliquary — Decentralised post-training on demand

Subnet History

SubnetIdentitySet 2026-08-15 11:52 UTC

Block 8,849,830Extrinsic

SubnetIdentitySet 2026-08-14 23:14 UTC

Block 8,846,041Extrinsic

SubnetIdentitySet 2026-04-10 02:54 UTC

Block 7,934,860Extrinsic

SubnetIdentitySet 2025-09-22 09:52 UTC

Block 6,498,357Extrinsic

SubnetIdentitySet 2025-08-17 09:26 UTC

Block 6,239,028Extrinsic

Recent Activity

SubnetIdentitySet Block 8,849,830 Extrinsic2026-08-15 11:52 UTC
SubnetIdentitySet Block 8,846,041 Extrinsic2026-08-14 23:14 UTC
SubnetIdentitySet Block 7,934,860 Extrinsic2026-04-10 02:54 UTC
SubnetIdentitySet Block 6,498,357 Extrinsic2025-09-22 09:52 UTC
SubnetIdentitySet Block 6,239,028 Extrinsic2025-08-17 09:26 UTC

This information was auto-extracted from the GitHub repository and may be incomplete or out of date. Always refer to the subnet's GitHub repository (github.com, opens in a new tab, external site) for the latest instructions.

Hardware Requirements

Miner

GPURequired

Not yet verified: GPU model, VRAM, Notes

Validator

GPURequired

Not yet verified: GPU model, VRAM, GPU count, Notes

Machine-extracted from the subnet's repository docs; may be wrong or out of date. Check the repository before buying or renting hardware. Open the repository (github.com, opens in a new tab, external site) · Extracted 2 Nov 2026

Miner Setup

Pull the pinned miner image (V1.0)

docker pull ghcr.io/reliquadotai/reliquary-validator:sha-3b3af1d-logic@sha256:7bb847499605308b1ad7b21f5430b1662eae506deacb6899d682bc0c11ce2742

Set environment variables for protocol v1 miner

export RELIQUARY_PROTOCOL_PROFILE=qwen3-4b-base-dapo-reliquary-v1 && export RELIQUARY_EXPERIMENTAL_FILL_CLOSED_ENABLED=1 && export RELIQUARY_TRAINING_RUN_ID=qwen3-4b-base-dapo-reliquary-v1-20260910

Start the miner with Math, Code and/or Logic environments

reliquary mine --network finney --netuid 81 --wallet-name YOUR_EXISTING_WALLET --hotkey YOUR_REGISTERED_HOTKEY --wallet-path YOUR_EXISTING_WALLET_DIRECTORY --validator-url http://62.238.81.36:8000 --environments openmathinstruct,opencodeinstruct,reliquary_logic_v2

Verify validator readiness before starting(optional)

curl --fail --silent --show-error --max-time 10 http://62.238.81.36:8000/readyz && curl --fail --silent --show-error --max-time 10 http://62.238.81.36:8000/state

Validator Setup

Export the job contract for a specific evaluation task

reliquary tasks contract --task-id <job> > <job>.contract.json

Set task environment variables

export RELIQUARY_TASK_ID=<job> && export RELIQUARY_TASK_CONTRACT=$PWD/<job>.contract.json

Start the validator process for a specific task

reliquary validate --wallet-name <wallet> --hotkey <hotkey> --http-host 0.0.0.0 --http-port <port> --no-set-weights

Qualify proof capacity with measured end-to-end latencies (production v3+)(optional)

python scripts/qualify_proof_capacity.py staging-proofs.jsonl --output proof-capacity.json --software-revision <full-release-commit> --checkpoint-revision <full-benchmark-checkpoint-commit> --runtime-fingerprint-hash <health-runtime-profile-hash> --hardware-class 'NVIDIA H100 80GB HBM3' --benchmark-device-count <count> --measured-at 2026-07-31T00:00:00Z

Publish a stamped base checkpoint with lineage (production v3+)(optional)

python scripts/publish_base_reset_checkpoint.py

No hardware? Back this subnet by staking instead New to mining? Start with the setup guide Validator setup guide

Day (UTC)

2026-11-04

0.0001%

2026-11-03

0.0001%

2026-11-02

0.0021%

2026-11-01

0.0027%

2026-10-31

0.0016%

2026-10-30

0.0038%

2026-10-29

0.02%

2026-10-28

0.0032%

2026-10-27

0.0022%

2026-10-26

0.02%

2026-10-25

0.02%

2026-10-24

0.01%

2026-10-23

0.0031%

2026-10-22

0.00%

2026-10-21

0.0003%

2026-10-20

0.0003%

2026-10-19

0.00%

2026-10-18

0.00%

2026-10-17

0.0008%

2026-10-16

0.0006%

2026-10-15

0.00%

2026-10-14

0.0004%

2026-10-13

0.0016%

2026-10-12

0.0022%

2026-10-11

0.01%

2026-10-10

0.01%

2026-10-09

0.01%

2026-10-08

0.0029%

2026-10-07

0.0012%

2026-10-06

0.0027%

2026-10-05

0.0006%

2026-10-04

0.0006%

2026-10-03

0.0015%

2026-10-02

0.0021%

2026-10-01

0.0006%

2026-09-30

0.0002%

2026-09-29

0.0002%

2026-09-28

0.0004%

2026-09-27

0.0003%

2026-09-26

0.0002%

2026-09-25

0.0003%

2026-09-24

0.0001%

2026-09-23

0.0001%

2026-09-22

0.0001%

2026-09-21

0.0003%

2026-09-20

0.0008%

2026-09-19

0.0023%

2026-09-18

0.01%

2026-09-17

0.03%

2026-09-16

0.11%

2026-09-15

0.23%

2026-09-14

0.25%

2026-09-13

0.22%

2026-09-12

0.20%

2026-09-11

0.10%

2026-09-10

0.18%

2026-09-09

0.35%

2026-09-08

0.35%

2026-09-07

0.33%

2026-09-06

0.37%

2026-09-05

0.42%

2026-09-04

0.42%

2026-09-03

0.43%

2026-09-02

0.39%

2026-09-01

0.31%

2026-08-31

0.31%

2026-08-30

0.31%

2026-08-29

0.33%

2026-08-28

0.34%

2026-08-27

0.33%

2026-08-26

0.32%

2026-08-25

0.32%

2026-08-24

0.29%

2026-08-23

0.27%

2026-08-22

0.35%

2026-08-21

0.42%

2026-08-20

0.43%

2026-08-19

0.43%

2026-08-18

0.54%

2026-08-17

0.59%

2026-08-16

0.58%

2026-08-15

0.56%

2026-08-14

0.58%

2026-08-13

0.57%

2026-08-12

0.59%

2026-08-11

0.62%

2026-08-10

0.62%

2026-08-09

0.60%

2026-08-08

0.57%

2026-08-07

0.55%

Reliquary implements this on Bittensor subnet 81.

reliquary mine --network finney --netuid 81 --wallet-name YOUR_EXISTING_WALLET --hotkey YOUR_REGISTERED_HOTKEY --wallet-path YOUR_EXISTING_WALLET_DIRECTORY --validator-url http://62.238.81.36:8000 --environments openmathinstruct,opencodeinstruct,reliquary_logic_v2

python scripts/qualify_proof_capacity.py staging-proofs.jsonl --output proof-capacity.json --software-revision <full-release-commit> --checkpoint-revision <full-benchmark-checkpoint-commit> --runtime-fingerprint-hash <health-runtime-profile-hash> --hardware-class 'NVIDIA H100 80GB HBM3' --benchmark-device-count <count> --measured-at 2026-07-31T00:00:00Z

What moved the price

Overview

Today so far

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Registration Activity

Development Activity

What Reliquary does

About

Website

Subnet History

Recent Activity

Participate

Hardware Requirements

Miner Setup

Validator Setup

What moved the price

Today so far

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

What Reliquary does

About

Website

Subnet History

Hardware Requirements

Miner Setup

Validator Setup