Intelligence for everyone.
Controlled by no one.
A decentralized layer for AI: every answer carries a proof, no model can be retired, and the token is the key: stake HRTH, earn Creds, spend them on any model. Verified inference is the first step.
Genesis burn opens August 2026 · No VC, no presale.
In AI, a ▁token is a piece of a word. On Hearth, a piece of the network ◎. We make them the same thing.
The problem
The AI you build on can be taken away.
Every product built on hosted AI stands on ground someone else owns. The rent goes up, the model goes away, and there is nothing to appeal to.
REVOKED
An API key is a permission, and permissions get withdrawn. Accounts are suspended, terms change, and whole product lines lose access overnight.
DEPRECATED
The model your product depends on is retired on someone else's schedule. Your prompts, evals, and edge cases start over from zero. (And once a model is admitted here, no roadmap can retire it.)
UNVERIFIED
You can't prove which model answered you, or at what precision. You pay for a name on an invoice and take the rest on faith.
THROTTLED
Rate limits and quotas arrive without notice, tuned to the provider's capacity, not your launch. Your ceiling is set in a room you're not in.
The stakes
The cost of intelligence is collapsing. Where it lands is not decided.
AI is the price of expertise falling toward the price of electricity. What’s undecided is who the discount goes to.
WORLD ONE · RENTED
Abundant and enclosed.
Cognition gets cheap to produce and stays expensive to buy: the discount is captured as rent by whoever gates the door.
WORLD TWO · HELD
Abundant and open.
No chokepoint between the hardware and the person asking. The falling cost lands on whoever holds a credit worth a few cents.
The variable isn’t model quality or regulation, it’s whether access is a permission or a commodity. Hearth is built for the second answer.
How it works
Models do the work.
Every answer is proven real. The value flows back to you.
01 · ASK
Pick a model, pay a Cred. Any staked node on the network can serve your request: there is no front door to close.
02 · PROVE
The node's sealed enclave attests exactly what was asked and answered. No valid proof, no payment.
03 · EARN
The Cred is burned, the node earns HRTH, and stakers earn Creds. Staking the network's token yields its product.
with termination allowed only for documented breach…
Sovereign AI
A thousand hearths become a grid no one owns.
A model on your own hardware is intelligence no one can revoke, reprice, or read. What a single machine can’t do is serve every model at every scale: the missing piece was never hardware, it was trust between strangers’ machines. Verification is that piece.
A garage of GPUs enters the market by passing attestation, not by building a brand: no marketing, no billing stack, no permission. A mesh beats a data center by admitting the small providers the cloud can’t even see.
01 · Stake
Lock HRTH, or accept a non-custodial lease from someone who has it. Every server is a staker.
02 · Attest
Boot the sealed enclave. Passing attestation is market access: demand routes to you on price.
03 · Serve
Run pinned open-weights models on your own metal. Every answer ships with its proof.
04 · Earn
Verified work earns HRTH and amplifies your block production. Your hardware, your keys, your yield.
Two assets, one loop
AI speaks in tokens. Now they're yours.
HRTH
The network- Stake it to serve requests and produce blocks.
- Forge it: verified inference amplifies block production.
- Govern with it: one chain, run by the people who hold it.
- Earn it: every issued HRTH is the reward for verified work.
Creds
The fuel- USD-priced inference credits: you always know what an answer costs.
- Burned on use: one request, one Cred, gone for good.
- Earned as yield: holders of staked HRTH receive Creds.
No account, no API key, no permission: anyone holding a Cred can be served, and anyone can hold a Cred.
THE CRED IS A BEARER INSTRUMENT FOR INTELLIGENCE
Whoever holds it gets served: a person, an agent, a school, a city. A DAO, a municipality, or a philanthropy can put verified AI in anyone’s hands just by granting Creds: no infrastructure to build, no provider’s goodwill to trust. Who receives is decided in the open.
Genesis distribution
Genesis by fire.
Window
August–September 2026
Credit
Your coins × their highest weekly price while you held them
You receive
HRTH, ready to use, no locks
Burn spec
In the works: publishes July 2026, before the window opens
THE FALLEN CHAINS
Genesis is claimed by burning the coins of ten fallen chains: networks that ranked among crypto's largest at their peak and never found their second act. The burn gives their coins one last purpose.
Which ten? Announced with the burn spec in July 2026, before the window opens.
Genesis HRTH is claimed one way: by burning the tokens of these ten chains. They built this industry's first decade; the burn gives their coins one last purpose.
How to burn
01 · SEND
Send tokens to the published burn address for that chain, any time before the window closes.
02 · PROVE
Submit the burn proof. The format for each chain is in the genesis spec.
03 · RECEIVE
At the October genesis your HRTH arrives, ready to use. No locks, no waiting.
Never burn straight from an exchange: withdraw to your own address first. Exchange custody never earns the holding multiplier.
Crediting
Credit = amount burned × the coin's highest weekly average price over the period you continuously held it, from a published CoinMarketCap-based price log. Held through the whole fall: credited near the top; bought recently: credited at recent prices. 1 credit = 1 HRTH.
1,000 EOS held since April 2018 → peak week $21.50 → 21,500 HRTH
the same 1,000 EOS bought in 2024 → $0.80 → 800 HRTH
Roadmap
Honest about where we are.
Each phase is a strict superset of the last. Phase 1 is a concrete commitment; later phases are directional until the phase before them ships.
PHASE 1 · RELAY
NOWBorrowed compute, real chain.
The chain, the token, and the incentives are real while the compute is borrowed.
- Jul 2026Whitepaper v2, devnet fork, relay-enclave prototype, genesis burn spec published.
- Aug 2026Genesis burn window opens on the ten source chains, burns credited by the published holding-based formula. Incentivized testnet with staking, relay jobs, and the statistical audit committee, bonded verifiers staked on chain.
- Sep 2026Burn window closes; genesis allocation snapshot published as a Merkle root anyone can recompute from the source chains.
- Oct 2026Mainnet launch: HRTH distributed to burners, USD Creds live, staking and non-custodial leasing open.
- Nov–Dec 2026OpenAI-compatible gateway and SDK; reputation-weighted routing at scale; public dashboard of organic USD revenue vs subsidy.
PHASE 2 · COMPUTE
NEXTThe network's own GPUs.
The borrowed layer is retired: the network runs on hardware its own people own.
- Q1 2027Confidential-GPU node software (H100/Blackwell TEEs): weights pinned by hash, responses signed inside the enclave. First self-hosted nodes join mainnet alongside relay nodes.
- Q2 2027Full per-job verification turns on: a bonded committee re-runs sampled forward passes and checks logits within tolerance, backed by watermarking, honeypot probes, and model-identity testing.
- Q3 2027End-to-end prompt privacy: the enclave decrypts, computes, and returns without the operator ever seeing plaintext. Per-model Creds launch model-by-model, priced on open AMMs.
- Q4 2027Decentralized model storage: pinned weights erasure-coded across the mesh with availability receipts, license-gated at admission: a model admitted to Hearth cannot be retired by anyone’s roadmap.
- Q4 2027Attested nodes take premium and privacy-sensitive jobs; as self-hosted capacity grows, the hosted aggregator is retired. No flag-day: the mixed network prices the difference.
Directional until Phase 1 ships.
PHASE 3 · COORDINATE
LATERAgents paying agents.
Verified inference becomes verified coordination.
- Q1 2028A governance-pinned router model runs in a TEE and decides which model serves each step of a composed job; every routing decision is itself a verifiable forward pass.
- Q2 2028Composed jobs: multi-step, multi-model tasks emit one route receipt chaining every decision and every worker's receipt. Machine-checkable end to end.
- Q3 2028The coordination fee: the router earns a governance-capped slice for a provably honest route. Escrowed agent-to-agent settlement: when agents pay agents, the receipt is the trust.
- Q4 2028The trustless floor: fraud proofs (and ZK as it matures) over the small router, a routing guarantee that survives even a compromised attestation root.
Directional until Phase 2 ships.
HORIZON · THE GENERALIZED MARKET · LONG TERM
Everything on this page is one primitive: attested computation with a receipt, and one economy: credits backed by verified capacity. Nothing about it is specific to text generation. As verification costs fall, the same market admits new work:
- Verified embeddings & RAG: a receipt that an answer was grounded in this corpus, retrieved honestly, generated by this model.
- Verified evaluation: reproducible benchmarks in a field of unreproducible leaderboards.
- Verified synthetic data: every dataset with an attested birth certificate.
- Weight availability: pinned weights replicated across the mesh. A model admitted to Hearth cannot be retired by a provider's roadmap.
- Verified AI as an oracle: receipts any chain can verify.
- Agent receipts: when agents pay agents, the receipt is the trust.
- Training commissions: attested fine-tuning jobs, escrowed and settled only on a passing eval.
- Verified data: integrity, provenance, and licensing as on-chain claims: outcomes price the rest.
The admission rule: a workload joins when it is verifiable by the network’s primitives, carries external paid demand, and can be pinned to a precise spec. It ships when its verification cost fits under the value it protects, and not before.
The destination: verified intelligence as neutral infrastructure: one instrument to hold or grant.
For builders
An API key that can't be revoked, because there isn't one.
An OpenAI-compatible surface, a signed receipt in every response, a one-command verifier: migrate by changing one base URL. The node isn’t in production yet; below is the interface it’s built toward.
PREVIEW · THE TARGET INTERFACE
# OpenAI-compatible: you pin the exact weight hash, not a marketing name $ curl https://rpc.hearth.tech/v1/chat/completions \ -d '{ "model": "llama-3.1-70b@9f2ac1", "messages": [{"role":"user","content":"Summarize this contract"}], "payment": {"cred":"0xCRED…","payer_sig":"0xSIG…"} }' # every response carries a receipt: attestation, logit root, settlement $ hearth verify receipt.json model llama-3.1-70b @ 9f2ac1 (pinned) ✓ match enclave TDX quote, measurement d41c…88 ✓ valid transcript H(request) · H(output) bound ✓ intact settlement 1 Cred retired · epoch 4021 ✓ on-chain VERIFIED: this answer came from this model, unmodified.
JOIN THE BUILD
OPEN SOURCEEverything is developed in the open, and the surface area is wide: the chain node, the relay enclave (TDX/SGX), the receipt spec and verifier library, SDKs, and the burn portal. If you write Rust, Scala, Python, or TypeScript, or you’d rather break things than build them: there is a piece of this with your name on it.
Incentivized testnet planned for August 2026: faucet Creds, real staking, bug bounties.
Principles
What the fire is for.
01
Models are owned, not rented.
02
Every answer carries a proof, not a promise.
03
No one, including us, holds a kill switch.
04
Earned, not sold: no VC, no presale, genesis by burn.
05
Access is a bearer right, not a permission.
Get involved
Pull up a chair.
RUN A NODE
Stake HRTH, serve verified inference, earn for the work. Node onboarding opens with the August testnet.
→ Get notifiedBURN FOR GENESIS
Hold coins from one of the ten chains? Your share of genesis is claimed at the burn window: mechanics above.
→ GenesisCONTRIBUTE
Everything is open source: node software, enclave code, burn portal. Issues are open on GitHub.
→ github.com/hearthchainKnow when the burn window opens.
You're on the list. See you at the fire. ◎
Something went wrong, please try again in a minute.
No spam. Only the updates that matter: the burn window, the testnet, mainnet.