Run neural networks inside smart contracts让神经网络直接跑在智能合约里

1 · the idea1 · 理念与设计

A model you can audit, because it is only integers只由整数组成，所以能被审计的模型

Most AI runs on a server you have to trust. A contract cannot work that way: it has no floating point and no network, and everything it computes must come out the same on every node. BNNX takes that as its design rule. A network is trained in PyTorch, then turned into integers: weights of +1 or −1, thresholds, small integer levels. Those integers are the model, and a reference program and a contract compute exactly the same ones. 大多数 AI 跑在你只能选择相信的服务器上。合约做不到这样：它没有浮点数，也没有网络，算出的每一个结果在每个节点上都必须相同。BNNX 把这当成设计规则：网络在 PyTorch 里训练，然后变成整数：+1 或 −1 的权重、阈值、几位的整数电平。这些整数就是模型本身，参考程序和合约算出的是完全相同的整数。

// ask a model from any contract
interface IBnnxModel {
    function classify(bytes calldata pixels)
        external view
        returns (uint256 label, int256[] memory scores);

    function modelInfo()
        external view
        returns (bytes32 name, bytes32 irHash,
                 uint256 inputLength, uint256 numClasses);
}

contract DigitReader {
    IBnnxModel public immutable model;

    constructor(address runner) {
        model = IBnnxModel(runner);
    }

    function read(bytes calldata pixels)
        external view returns (uint256 label)
    {
        (label, ) = model.classify(pixels);
    }
}

Where BNNX fits in TapeOutBNNX 在 TapeOut 里的位置

TapeOut already puts a site's front end on chain (DeWeb) and its state in a container. What a site cannot yet do is call logic that is just as verifiable. BNNX is the project that supplies that missing layer: a trained model becomes a pure contract that a page, a circuit's owner or another contract can call, and anyone can check. TapeOut 已经把网站的前端放上了链（DeWeb），把状态放进了容器。网站目前还做不到的，是调用同样可验证的逻辑。BNNX 就是补上这一层的项目：训练好的模型变成一个纯函数合约，页面、电路持有人或别的合约都能调用，任何人都能核对。

2 · where it stands2 · 项目进展

From a first digit model to a whole toolchain从第一个数字模型到一整套工具链

Everything below was built and checked on BSC testnet between 2026-09-28 and 2026-10-22. Each line points at the report it comes from. 下面的一切都在 2026-09-28 到 2026-10-22 之间做完，并在 BSC 测试网上检查过。每一条都指向它所依据的报告。

1. 2026-09-28

   The first two models on chain前两个模型上链

   Binarised networks for digits (97.86%) and small photos (81.42%). The chain returns exactly the integers the reference runtime computes.二值化网络，识别手写数字（97.86%）和小图（81.42%）。链上返回的整数和参考运行时算出的完全一样。

2. 2026-09-30

   The developer toolkit开发者工具包

   A Python package and the bnnx command, training layers, a JavaScript SDK, a Solidity interface, receipts, a model zoo and a tutorial.Python 包和 bnnx 命令、训练层、JavaScript SDK、Solidity 接口、收据、模型库和教程。

3. 2026-10-06

   Graph models and shared operators图模型和共用算子

   A model becomes a graph of nodes; small pure operator contracts are deployed once and shared by every model.模型变成由节点组成的图；小的纯算子合约只部署一次，所有模型共用。

4. 2026-10-08

   Board layers, and a gomoku network棋盘类的层，和一个五子棋网络

   Padding, 1×1 convolutions, lookup tables and adds. A contract calls the network in one transaction and gets a correct move.补边、1×1 卷积、查表和相加。合约在一笔交易里调用网络，得到正确的一步棋。

5. 2026-10-18

   Few-bit layers多位的层

   4-bit activations and int8 weights. A CIFAR-10 network (79.63%) that a contract can call inside a single transaction.4 位激活和 int8 权重。一个 CIFAR-10 网络（79.63%），合约在单笔交易里就能调用。

6. 2026-10-19

   Networks trained by someone else别人训练的网络

   The QONNX importer takes a Brevitas network on chain. On 10,000 test images its labels match the original's own executor exactly.QONNX 导入器把一个 Brevitas 网络搬上链。在 10,000 张测试图上，标签和原模型自己的执行器完全相同。

7. 2026-10-20

   Bigger models更大的模型

   A new dispatcher takes up to 64 data chunks. A 21-chunk digit network is the first model past the old limit of 16.新的调度合约最多支持 64 个数据块。一个 21 块的数字网络，是第一个超过原来 16 块上限的模型。

8. 2026-10-21

   A second imported network第二个导入的网络

   LFC from the QONNX model zoo: 98.81% on MNIST, no label differs from the original on 10,000 test images.来自 QONNX 模型库的 LFC：MNIST 上 98.81%，10,000 张测试图上没有一个标签与原模型不同。

9. 2026-10-22

   A whole game, played on chain一整盘棋，在链上下完

   A game contract lets the network play inside the player's own transactions. One full game is on BSC testnet, and every AI move was re-checked on two nodes.对弈合约让网络在玩家自己的交易里落子。一整盘棋在 BSC 测试网上，AI 的每一手都在两个节点上复核过。

The toolkit工具包

What you could build with it可以拿它做什么

The first and last are done. The rest are ideas: nothing here is built yet.第一个和最后一个已经做到了，其余是想法：这里没有哪一个已经做出来。

3 · demos3 · 演示

Every box here is live这里每一个框都是活的

Each demo sends a free read-only call to contracts on BSC testnet, through two independent nodes. Nothing is signed. When two nodes answer differently, or one does not answer, the page says so. 每个演示都向 BSC 测试网上的合约发免费的只读调用，经过两个互相独立的节点。不需要签名。两个节点的回答不同，或者有一个没有回答，页面都会直接说出来。

4 · for developers4 · 给开发者

Train it, deploy it, call it训练，部署，调用

The toolchain is a developer preview: the repository is private while the licence is settled. The steps below are the ones we ran for every model on this page, and the commands are real. Calling and checking the deployed models needs none of it. 工具链现在是开发者预览：许可确定之前仓库是私有的。下面的步骤，是我们给这页上每个模型都走过的步骤，命令也是真实的。调用和核对已部署的模型完全用不到这些工具。

cast call 0x9B1FE44987bC449d28aC8c77e6027A6596c91901 \
  "modelInfo()(bytes32,bytes32,uint256,uint256)" \
  --rpc-url https://bsc-testnet-rpc.publicnode.com

# 0x64312d6d6e697374...   the name, "d1-mnist"
# 0x4fd07609d945...b831ec6d   the IR hash
# 784                     input bytes
# 10                      classes

import bnnx.nn as bn

model = bn.templates.mlp((28, 28, 1), hidden=(256,),
                         classes=10)
print(model.summary())  # size, gas per call, a verdict

bn.fit(model, (x_train, y_train), (x_val, y_val),
       epochs=5, batch=256, lr=1e-2, seed=0)

# checks the integer model against PyTorch,
# then writes it and returns its IR hash
ir_hash = model.cpu().export(
    "models/my-digits", name="my-digits",
    classes=[str(d) for d in range(10)])

# size and gas per call, before you spend anything
bnnx inspect models/my-digits

# dry run: simulates, sends nothing
bnnx deploy models/my-digits \
  --rpc-url https://bsc-testnet-rpc.publicnode.com \
  --account dev

# for real, signed by your own Foundry keystore
bnnx deploy models/my-digits \
  --rpc-url https://bsc-testnet-rpc.publicnode.com \
  --account dev --record my-digits.json --broadcast

# the chain against the model file and the reference
# program, on two nodes
bnnx verify models/my-digits \
  --deployment my-digits.json --random 100 \
  --rpc https://bsc-testnet-rpc.publicnode.com \
  --rpc https://bsc-testnet-dataseed.bnbchain.org

import { classifyOnChain, zooDeployment,
         zooModel, checkModelInfo } from "bnnx";

const nodes = [
  { label: "publicnode",
    url: "https://bsc-testnet-rpc.publicnode.com" },
  { label: "bnbchain",
    url: "https://bsc-testnet-dataseed.bnbchain.org" },
];
const { runner } = zooDeployment("d1-mnist", 97);

const identity = await checkModelInfo({
  address: runner,
  expectedIrHash: zooModel("d1-mnist").irHash,
  nodes,
});
const result = await classifyOnChain({
  address: runner, pixels: new Uint8Array(784), nodes,
});
console.log(identity.ok, result.agree,
            result.calls[0].label, result.calls[0].scores);

bnnx import-qonnx TFC_1W2A.onnx models/tfc-w1a2-qonnx \
  --classes 0,1,2,3,4,5,6,7,8,9

Limits you should know about需要知道的限制