This site lists neural networks whose weights and programs are stored fully on Ethereum mainnet. Each entry has a contract address and a live check that you can run from this page.

Models that store their files somewhere else, or run somewhere else and only post a proof, do not qualify. Each entry says where the model runs — inside Ethereum, or in your browser.

CENSUS

Registered onchain models
MODELAUTHORYEAR
陰XOR陽nahiko2021
INCHAINPEPEGANlifeonchain2023
CONCRETEHiggs × diid2025
ARTIFICIAL AFTER ALLHan Saglam2026
AUTOMATE ATTENTIONdiid2026
HELLO WORLD COMPUTERHan × 10012026
THE MARKET IS TALKINGFelixFelixFelix2026

7 entries. Each entry passed the automatic registry checks. Select an entry to run a new live check.

陰XOR陽

NAHIKO

陰XOR陽 facts
FIELDVALUE
ARCHITECTURE2-2-1 MLP
NEURONS3
TASKXOR
WEIGHTSfixed in contract storage
ACTIVATIONinteger sigmoid approximation
INPUTStwo onchain booleans
OUTPUTthe XOR result as one boolean
ADDRESS0x2A648cb23D3ADf1BB171Fabc0119eA3eb951841f
YEAR2021

Three neurons in two layers compute XOR. The artist minted this work in August 2021 as a one-token NFT. The artist and the collector each set one input. The network turns the two inputs into yin, yang, or a blend. The weights are fixed values in the contract; anyone can read them. isYingYang() runs the model fully inside Ethereum.

VERIFICATION

The check reads the contract code and runs the call on public Ethereum servers.

isYingYang() — runs the model on the two stored inputs and returns the XOR output as a bool

Requests go from your browser to public third-party Ethereum servers.

INCHAINPEPEGAN

LIFEONCHAIN

INCHAINPEPEGAN facts
FIELDVALUE
ARCHITECTUREDCGAN-style generator
PIPELINE16-dim latent → dense → 2× conv → 5×5 RGB
WEIGHTSint16 fixed-point, in contract storage
PARAMETERS≈2,900
ACTIVATION256-entry logistic lookup table
TRAININGoffchain — GAN trained on pepes, generator ported to Solidity
OUTPUTthree 5×5 frames, assembled onchain into an animated GIF
ADDRESS0x65e9A5BbC1B26C0833B232355534cC404b0e69D4
YEAR2023

This GAN generator draws animated pepes inside Ethereum. A call to tokenURI() takes three seeds from the token id, runs the network three times, and builds an animated 5×5 GIF from the three frames. The artist trained the weights offchain and wrote them into contract storage. The weights, the network, and the GIF encoder all live at one address. The image exists nowhere else.

VERIFICATION

The check reads the contract code and runs the call on public Ethereum servers.

tokenURI(1) — runs the generator three times (one per frame) and builds the animated GIF inside Ethereum

Requests go from your browser to public third-party Ethereum servers.

CONCRETE

HIGGS × DIID

CONCRETE facts
FIELDVALUE
ARCHITECTUREGAN generator, single upsampling block
PIPELINE16-dim noise → dense 4608 → 6×4×192 → batchnorm → ReLU → conv transpose
WEIGHTS79,680 int16 fixed-point, in SSTORE2 data contracts
ACTIVATIONReLU; rational tanh approximation on the output layer
INPUTSa 16-byte seed chosen by the collector at mint
OUTPUT12×8 RGB fingerprint, rendered as an onchain SVG
ADDRESS0x4512759606f9E680361DD76C4c17F92031De7AcD
YEAR2025

The first collection of Higgs, an AI artist that diid built. The contract holds the first stage of the GAN generator; its weights live in data contracts on Ethereum. fingerprintSvg() runs the network inside Ethereum on a 16-byte seed and returns a 12×8 RGB fingerprint as an SVG image. The large display images are offchain; the check here covers only the onchain generator.

VERIFICATION

The check reads the contract code and runs the call on public Ethereum servers.

fingerprintSvg(<token #1 seed>) — runs the model inside Ethereum; the call is too large for one public Ethereum server, so checks usually pass with 2 of 3 servers

Requests go from your browser to public third-party Ethereum servers.

ARTIFICIAL AFTER ALL

HAN SAGLAM

ARTIFICIAL AFTER ALL facts
FIELDVALUE
ARCHITECTUREcompressed decoder NN
PIPELINE32-dim latent → dense → 3× transposed conv → 1×1 conv
WEIGHTSint8-quantized, Huffman-compressed, in bytecode
OUTPUT32×32 grayscale, 1024 bytes
CONTRACTS256 sovereign pieces via factory
DEPENDENCIESnone
ADDRESS0xbDF5A7F5e5AE18eE580139297C63a89fc1d18B5f
YEAR2026

This work is 256 contracts on Ethereum. Each contract holds one seed and one compressed neural network in its own code. A call to artificial() runs the network inside Ethereum and returns a 32×32 grayscale image (1024 bytes). The contract computes the image at call time and stores no image. The address here is one example piece; a factory deployed all 256.

VERIFICATION

The check reads the contract code and runs the call on public Ethereum servers.

artificial() — returns a 32×32 image as 1024 pixel bytes (1088 raw bytes with ABI framing)

Requests go from your browser to public third-party Ethereum servers.

AUTOMATE ATTENTION

DIID

AUTOMATE ATTENTION facts
FIELDVALUE
ARCHITECTURE5-layer Llama-style transformer
DIMENSIONSdim 48, hidden 128, 8 heads, grouped-query attention
VOCABULARY2048 BPE tokens
WEIGHTSINT8 per-channel quantized, stored in onchain data contracts
INFERENCERoPE, RMSNorm, SiLU, KV cache — all in EVM opcodes
OUTPUTone aphorism per NFT, rendered as onchain SVG
ADDRESS0x5F8E7D750E75b44747C058A204D8DEa0D18fA5d3
YEAR2026

This contract is a complete transformer language model that runs inside Ethereum. Data contracts on Ethereum store its weights. A companion NFT contract calls the model to write one permanent aphorism for each token. The model also draws the text as an SVG image inside Ethereum. The address here is the language model itself.

VERIFICATION

The check reads the contract code and runs the call on public Ethereum servers.

logits([1]) — runs the model once and returns int16 scores (logits) for all 2048 vocabulary tokens

Requests go from your browser to public third-party Ethereum servers.

HELLO WORLD COMPUTER

HAN × 1001

HELLO WORLD COMPUTER facts
FIELDVALUE
ARCHITECTUREsingle-layer perceptron
FEATURES1024 hashed text n-grams
WEIGHTSint32 × 1024, packed in uint256[128] storage
TRAININGonchain — perceptron rule at question settlement
TASKbinary text classification
INITIAL STATEall weights zero
OUTPUTclass label with margin-based confidence
ADDRESS0x6901afceb66564a9b6E7193561fB78F0878b5906
YEAR2026

This perceptron lives inside one contract. The contract stores the weights, trains them, and answers questions. You send it a short text (up to 64 bytes). ask() returns a yes/no label and a confidence value. All weights started at zero. Anyone can post a yes/no question for a fee. Each settled question updates the weights inside the contract.

VERIFICATION

The check reads the contract code and runs the call on public Ethereum servers.

ask("hello world") — runs the model on the text and returns (label uint8, confidence int256)

Requests go from your browser to public third-party Ethereum servers.

THE MARKET IS TALKING

FELIXFELIXFELIX

THE MARKET IS TALKING facts
FIELDVALUE
ARCHITECTUREtwo-layer deep echo-state network
NEURONS20 hidden (two baskets of 10 ERC-20 markets), 4 inputs
WEIGHTSderived each block from market correlations, not stored
RENDERERengine.js + renderer.js on EthFS, served via scripty.sol
MATHfixed-point BigInt, integer tanh — bit-reproducible
OUTPUTone deterministic oracular sentence per block
ADDRESS0xd6D8555E131c0C431601aF62141A61a9b1aa6b13
YEAR2026

This network reads the market itself. Its hidden layers are two baskets of live token markets. Each block, it derives fresh weights from the trade activity of those tokens. Ethereum stores the renderer program; your browser runs it. The output is one sentence per block, from a 299-word vocabulary. stateAt() returns the exact state at any past block, so anyone can compute each sentence again.

VERIFICATION

The check reads the contract code and runs the call on public Ethereum servers.

stateAt(25599744) — returns the saved inputs and mood state; the renderer rebuilds the network from these

Requests go from your browser to public third-party Ethereum servers.

METHOD

Software checks the schema, file name, evidence links, and declared return size of each entry. Two or more public Ethereum servers must then agree on the contract code and the exact call result at one block. These checks prove that the call is reproducible. They do not prove that the code is the described neural network. A person also reviews the source evidence. To correct or add an entry, open a pull request.