Two small models walk in. One bigger model walks out.
Every MERGEBORN is an AI model drawn on-chain: a tier, an architecture and a parameter count. Train it once a day for EPOCH. Burn two of the same tier plus EPOCH to merge them into the next tier, which trains far more than the two did together. The first wallets to merge a Large or a Frontier are paid from the prize pot automatically.
Mint a genesis model
Your models
Train: each model once every 24h. Merge: tap two models of the same tier, then Merge. Fresh merges rest 24h before their first training.
Sell EPOCH to the redeem pool
EPOCH can always be sold back to its own contract for BNB. You receive amount × pool ÷ max(EPOCH supply, 200,000), minus 2% that stays in the pool. The pool only grows from paid mints and is never touched by anyone else: no owner, no withdraw function.
| Pool | … |
|---|---|
| EPOCH supply | … |
| 1 EPOCH → | … |
| Ever trained | … |
First to Large, first to Frontier
15% of every paid mint sits in the prize pot inside the NFT contract. When a merge creates a Large, the first 10 different wallets to do it are each sent 4% of the pot in that same transaction.
When a merge creates a Frontier, the first 5 different wallets are sent pot ÷ slots left: one fifth, then a quarter of what remains, and so on; the fifth Frontier empties the pot. After that, the prize share of new mints flows into the EPOCH pool instead. Nobody, including the deployer, can take the pot.
| Prize pot | … |
|---|---|
| Large prizes | … |
| Frontier prizes | … |
Methods
| Tier | At mint | EPOCH / day | Merge two (EPOCH) | Parameters |
|---|---|---|---|---|
| Nano | 68% | 10 | 10 → Mini | ~96M–144M |
| Mini | 24% | 24 | 30 → Base | ~480M–720M |
| Base | 8% | 60 | 80 → Large | ~2.4B–3.6B |
| Large | — | 150 | 200 → Frontier | ~11B–17B |
| Frontier | — | 400 | top tier | ~56B–84B |
Architecture (Transformer, Mixture-of-Experts, State-Space, Diffusion, Recurrent) is inherited from one parent at random; generation = older parent + 1. Randomness uses block data and is not a secure oracle. Models are standard ERC-721 and trade on any BNB Chain NFT marketplace.
Merging is the new training
This week's open-model news is about merging, not just bigger clusters. On Nov 4 a 4B open-weights decision model was fine-tuned across 16 internet-connected GPUs with staggered fragment merges instead of one data center. A day earlier a subnet team blended two miners' models with two open-weight models into a single release. Researchers are even editing weights directly.
MERGEBORN turns that loop into a game: many small models, trained in parallel by many wallets, merged step by step into a few large ones. It is a collectible and a game, not an investment, and it makes no promise about any price.
Sources: tao.media (Nov 4 2026) · subnetradar.com (Nov 3 2026) · codeandchain.com (Nov 3 2026)
Contracts (BNB Chain, verified on Sourcify)
The deployer address is shown as owner for marketplaces only; the contracts have no admin functions: no pause, no price change, no withdraw of the pot or the pool. Project share of mints goes to the FeeVault, which only its fixed owner can withdraw.