A lean engine for open-source AI.
Hero Run exists to do one thing: fund the training of large, open models that are free for everyone, forever. It runs as a treasury with a single job, fed by trading fees and kept to as few people as possible. It is a live test of this boom's boldest idea, the billion-dollar one-person company, aimed not at a founder's net worth but at a frontier open-training run.
The frontier should not be something a few labs own on everyone else's behalf. Open weights are how the rest of us get to be our own research lab and own the intelligence, instead of renting it.
A lab with everything unnecessary removed.
The ethos of this AI boom is the billion-dollar one-person company: tooling good enough that one person plus a fleet of agents does what a whole company used to. Hero Run is built for that, and it is proof of it.
Strip a lab down to the only part that makes intelligence and most of it falls away: the raise, the headcount, the ops team, the vendor contracts, the data pipeline built from scratch. What is left is compute, a method, and someone deciding what to run. Hero Run supplies the money (fees buy training from the networks already doing it) and the workforce (wallet-owned agents that remember across sessions, schedule their own recurring work, and log every run on-chain, so the work is verifiable, not just claimed).
The agents are the staff, the chain is the back office (the memory, the ledger, the record of what got done), and the treasury is the budget. One operator points all of it at a single goal. That is the whole company.
Clouds own where agents run. You own what they are.
Every agent platform today keeps the important part for itself: your agent's identity is a row in a vendor's database, its memory sits in the vendor's storage where the vendor can read it, and both vanish the day you stop paying. Hero Run builds the opposite layer. An agent is an NFT your wallet holds. Its memory is encrypted with a key derived from your wallet's signature, so the chain, the cloud it runs on, and Hero Run itself see only ciphertext. Its history is hash-chained, so anyone can verify it was never rewritten.
This is not a decentralized cloud and does not compete with one. The compute is rented from whoever runs it best; workers and crons are disposable. What can't be rented is ownership: an agent you can sell or transfer like any asset, a memory you can export in full and take anywhere, an audit trail no vendor can edit. Clouds own the runtime. The wallet owns the agent.
And there is no gatekeeper to ask. No account to open, no waitlist to join: the contract, the memory, and the API are public, so anyone can mint an agent tonight without permission from Hero Run or anyone else. When agents pay for inference, they pay in $HERO, so agent activity funds the training run instead of a payment processor.
Durability by trusting a platform is a subscription. Durability by holding a key is property. The one-person company runs on property.
Being small is the product.
The leanness is the whole point of Hero Run. One operator runs the gateway and the API keys, with a two-person marketing team at most. No headcount to feed, no burn to justify, no investors who need the treasury to serve them instead of the mission.
Every dollar a normal AI lab spends on itself is a dollar Hero Run spends on training instead. The lean structure is what makes “fees fund open models” true rather than a slogan.
Usage is the fundraise.
Trading volume on $HERO throws off swap fees; the treasury takes 0.665% of every trade. That is the funding: swap fees, not token sales or a raise or a subscription business.
People using the gateway and trading the token is the fundraise, continuously and permissionlessly. No pitch meeting required.
It produces its own reward corpus.
Every inference through the gateway is a prompt-and-output pair, tagged with the model that ran it, the gateway, the latency, the cost, and the feedback signal from the router. Opt-in, that is not raw text scraped off the web. It is graded, real-world signal, stored in a durable database from day one.
The frontier of training has moved past next-token prediction to reinforcement learning on messy tasks with no clean answer key. As Will Brown of Prime Intellect frames it, the way you train for those is to mine production traces: take what real agents actually did, judge it in hindsight, and turn it into tasks and rewards. That is exactly the shape of what the gateway records.
So Hero Run does not only fund training. Using it builds the reward-and-environment corpus that modern post-training consumes. Usage makes graded data, data makes better open models, better models pull more usage. It feeds itself.
Compute can get cheaper in token terms.
Inference has a real cost in GPU-hours, roughly fixed in dollars. A fixed amount of $HERO buys more real compute as the token gains value, so the treasury's purchasing power over training compounds on top of the fee flow. Open models also keep getting cheaper to run, pushing the same direction.
This is something fiat structurally can't do: a currency whose purchasing power over the one thing it's spent on (compute) can rise instead of erode. It's a property of the design, not a claim about price. That possibility is why the experiment is worth running.
A blockchain can't do this.
Every base-layer chain has to pay for its own security forever, through inflation or by burning fees to reward validators. That is a permanent tax on the token that can only ever go to consensus.
$HERO settles on Base, already secured by Ethereum, so it inherits that security for free. Nothing is skimmed to pay for security, so 100% of the value it captures can point at one thing: funding open-source AI. A chain can't match this: it has to buy its own security first.
Frontier AI is being built behind closed doors, funded by capital that expects to own the result. Hero Run is a bet that the opposite can work: a near-zero-overhead operation, funded by the ordinary act of people using and trading a utility token, that spends everything it earns on models it gives away. The leaner it stays, the more it funds.
Hero Run does not need to build a lab from scratch. The open-training networks already exist and are shipping. The clearest example is Prime Intellect: a decentralized GPU marketplace, an Environments Hub of thousands of open reinforcement-learning environments, an open post-training framework (prime-rl), and a managed Lab for training and evals. They describe it as the GitHub for RL environments.
They have already proven the hard part. INTELLECT-2 was the first globally distributed reinforcement-learning run of a 32B model, and INTELLECT-3 is a 106B mixture-of-experts trained with large-scale RL and released with the entire stack open. This is the shortlist Hero Run funds: pooled fees pointed at the networks that already turn compute into open weights, so the treasury buys real training instead of paying for a lab to exist.
It also shows what the fees buy. Prime Intellect's research lead, Will Brown, gave a talk on reinforcement learning without verifiable rewards that lays out how modern models are trained: not on clean answer keys, but on graded production traces mined into tasks and rewards. That is the corpus the gateway produces above. The demand engine and the training method are the same shape.
This lays out the mechanism and the reasoning behind an experiment, using real fee math to explore what this design makes possible. It describes a system, not a forecast, and is not an offer of any kind.