Overview
Inference Markets leverages a new primitive: Intelligence. Every market launched here is quoted in a unit of model output instead of a stablecoin, its holders are paid the trading fees in that unit every block, and the unit itself can be burned for real inference on Venice AI. It runs on Base, it is powered by Venice AI, and the chain is its only backend.
What the primitive means
Every market is built on a primitive: the chain's gas token, a stablecoin, or nothing at all. Here it is intelligence, and it pays you in model output.
When you launch a token on Inference Markets you pair it with a unit of intelligence rather than with dollars — for example INT, which is one million output tokens from a frontier model. Your token's price is quoted in INT. When people trade it, the trading fees are collected in INT and accrue to the people holding your token, block by block. Claim them whenever you like. If you hold the token, you are being paid in model output.
INT has a real price, and nobody sets it by hand. Once an hour a Chainlink Functions network reads what the top models actually charge per million tokens and writes the sample on-chain. The contract keeps a seven-day median, and the protocol keeps the coin trading at that price with its own liquidity. So one INT is always worth what a million frontier tokens cost today.
You can do three things with the INT you are paid. Keep it and hold the price of intelligence. Burn it for dollars at the index, any time, with no spread. Or burn it for DIEM, Venice AI's compute token: one DIEM staked in your Venice account gives you a dollar of Venice API credit every day, for as long as you hold it. A payout you received once becomes inference you can use from then on.
Or skip the coin entirely and tell the protocol what you want to be paid in. Dollars. DIEM. NVIDIA. Your fees are converted when you claim them, before they reach your wallet.
The chain is the backend
There is no server between you and any of this. Stock and ETH prices are read from Chainlink at the moment a contract needs them. DIEM and VVV are priced from Aerodrome's on-chain six-hour averages. The inference index is written by a Chainlink Functions network from OpenRouter's public price list, and its model lists live on-chain behind a 24-hour delay. Your fees accrue inside the token you hold, and a claim is a transaction you send yourself. Repricing a peg, collecting fees, sweeping the protocol's share, topping up the vault and keeping its DIEM reserve in band are public functions anyone can call, each with a floor taken from the on-chain feeds, so a stranger calling them cannot make the protocol trade below the index. Nobody holds a key that can set a price or pay a holder. Automation lists what has to be called and by whom.
The three parts
Inference coins are ERC-20 tokens, 18 decimals, one per unit: INT is a million frontier output tokens, OPEN a billion open-weight output tokens, AGENT an hour of an agent working. Each of the three is held at its reference price by a single-sided Uniswap V3 pool that anyone can reprice as the feed moves. DIEM and six Coinbase tokenized AI stocks are pair coins in their own right, routed through Aerodrome, which makes ten pair coins in all.
Markets are Uniswap v4 pools, launched from a single form in a single transaction. A market pairs a new token of fixed 1,000,000,000 supply against the coin the creator picked, opens at a $5,000 market cap, and holds its whole supply as concentrated liquidity from the block that creates it. Both positions belong to the protocol, which has no function that withdraws them.
Rewards accrue every block. Every swap pays an LP fee of 1% to 3%, fixed by the creator at launch. When the fees are collected, the coin side is split 50% to the token's holders, 30% to buy and burn the protocol token, 20% to the protocol, and the token side goes to holders in full. The holders' share is credited to the token's own per-share accumulator, so every wallet's balance earns for exactly the blocks it was held. A holder claims in the pair coin, or in whatever asset they asked to be paid in instead.
What backs an inference coin
The three synthetic coins are minted by the protocol into the ask side of a V3 pool at the feed price. The dollars buyers pay in are forwarded to the RedemptionVault, which pays them back at the index whenever someone burns a coin, in USDC or in DIEM, in the same transaction. Part of each reserve is held as DIEM so that delivery is a transfer rather than a trade.
Coverage is what the vault holds divided by what it owes, published on-chain per coin and targeting 100%. Redemption and Coverage cover both, and Risks says plainly where this can go wrong.
Chain
Inference Markets is deployed on Base, chain ID 8453, an OP-stack L2 that uses ETH for gas. The USD stable is Circle's native USDC. DIEM, VVV, the tokenized stocks, Aerodrome, Chainlink, and canonical Uniswap V3 and v4 are all on Base, which is why a redemption is one transaction. Addresses are in Contracts.