Brief · v1

Liquid Tensor

A decentralized machine-learning network where holding the token runs a Subnet Node that hosts AI model shards and mines lTAO. This is how the pieces work, on Uniswap v4, fully on-chain.

00Abstract

The future of AI should not be owned by a handful of data centers. Liquid Tensor is a peer-to-peer network for machine learning: GPU peers each host a piece of the models and combine into subnets that serve them, the same way BitTorrent shares a file. $lTAO is the token that runs it. Hold $lTAO and a Subnet Node mints to your wallet, represents your slice of the network, and mines lTAO from a Bitcoin-style emission. Nothing is staked to start, there is no mint page, and the art is a subnet drawn entirely on-chain.

01The network

Liquid Tensor is built on proven decentralized-learning research (Hivemind over libp2p, with a Kademlia distributed hash table to connect many machines into one network). Per model, each peer hosts a shard and combines with the others, so anyone can train, fine-tune, and sample models the way they would on PyTorch or Hugging Face.

  • Subnets are decentralized p2p networks that each serve AI models (language, vision, diffusion and more), like a blockchain that computes models instead of transaction blocks.
  • Inference validation is optimistic: node validators check that other nodes are honestly hosting the models. A dishonest node is removed by consensus, with penalties including slashing.

02The hybrid (how the node works)

$lTAO is a DN404-style hybrid. The ERC-20 and the Subnet Node ERC-721 are linked: your node count always equals your balance divided by 600. Cross 600 and a node appears, cross 1,200 and a second one appears, sell below a line and the matching node returns. There is no separate transaction to mint, it happens inside the transfer.

  • The pool, the hook and other contracts are flagged skipNFT, so the pool never mints nodes. Only real holders do.
  • Because a node mints per whole 600 and a buy is capped at 2,100, a single buy mints only a few nodes, which keeps buying cheap. The 600 figure is 20% of 3,000, sized to the 20% of supply that actually sits in the pool.

03Halving emission

This is the heart of the tokenomics, and it mirrors Bitcoin. Of the 21,000,000 fixed supply, only 20% seeds the pool. The other 80% is an emission reserve held by the token itself, mined out to Subnet Nodes over 64 halvings. There is no proof-of-work, the network rewards compute instead.

  • The reserve emits at a base rate that halves every year, so early nodes mine the most, exactly like early Bitcoin blocks.
  • Each second of emission is split across all nodes by weight, using a continuous magnified accumulator, so the maths is exact to the wei.
  • You call claimReward() to collect your mined $lTAO at any time, and claiming can compound straight into more nodes.

04Stake & tenure

Following nominated proof-of-stake, a node can be staked to run as a validator. Staking locks the node's 600 backing and makes the node permanent, then starts a tenure that grows its weight the longer it stays staked.

  • A live node has a reward weight of 100. A staked node's weight climbs to 250 over 90 days, so a fully tenured node mines 2.5x the lTAO of a fresh one.
  • Unstaking returns the backing minus an 8% burn, the on-chain echo of a slash, and retires the node.

05Fee & the pool

$lTAO trades against ETH in a Uniswap v4 pool. A v4 hook takes a 4% fee on the ETH side of each swap and sends it to the treasury behind the network. The hook does no extra work on a swap, so it does not add cost to a buy beyond the small fee, and the pool fee itself is zero.

06Limits & launch

The launch is fair: the 20% liquidity tranche was paired into the pool with 1 ETH, with no team allocation. Two limits are enforced on-chain and switch on the moment trading opens:

  • Max 2,100 $lTAO per buy, so no single transaction can sweep the book.
  • Max 42,000 $lTAO per wallet, so no one wallet can corner the early float.

07On-chain art

Every Subnet Node is a subnet graph drawn on Ethereum, not hosted on a server, and split across roughly 56 small verifiable contracts behind one composer: 32 per-subnet generators, palette banks, and base shards. The result is animated, with peers and edges, signals travelling the mesh, and a glowing frame once a node is staked. Because it is read from the chain, a node visibly changes as you hold and stake it.

08Parameters & contracts

Supply21,000,000 $lTAO, fixed
Emission reserve80% mined to nodes over 64 halvings
$lTAO per node600
Emissionhalves every year, by stake-weight
Stake tenure1.0x → 2.5x over 90 days
Unstake burn8% of the backing
Max buy / wallet2,100 / 42,000
Swap fee4% (the v4 hook)

Contracts: Liquid Tensor ($lTAO + the hybrid + emission + stake), SubnetNode (the ERC-721), LiquidTensorHook (the v4 hook), SubnetRenderer (one contract that deploys its own art modules). Hold $lTAO, run a node, mine lTAO. Everything above runs on-chain the moment you cross 600.