BNNX

Preview copy · BSC testnet, not mainnet. Read-only calls: no wallet, nothing to sign.预览副本 · BSC 测试网，不是主网。只读调用：不需要钱包，没有签名。

BNNX v0 · BSC testnet · chain 97

Neural networks that answer inside one eth_call

Two binarized networks run as contracts. Each classification on this page is a free, read-only call sent to two independent RPC operators. The answers are integers, so the two nodes either agree exactly or they don't.

Handwritten digits

Draw one digit, large and centred
What the network sees: 784 bits, one per pixel ≥ 128

Your drawing is cropped, scaled and centred to 28 × 28 grey pixels in your browser; only those 784 bytes are sent, as call data, to the two RPC nodes. The contract's first layer turns them into the bits shown above.

Small colour photos

Pick a test photo or upload your own

CIFAR-10 test photos

Uploads are centre-cropped and shrunk to 32 × 32 in your browser; only those 3,072 bytes are sent, as call data, to the two RPC nodes.

What you are calling

Check it yourself

Read the model's identity straight from the chain with Foundry. The second value is the IR hash: the SHA-256 that pins the model's architecture and weights.

How the deployments were verified (read-only, 2026-09-28, on both nodes): the runner's code was compared byte for byte with the build of its Solidity source, every weight chunk with the model files the IR hash names, and 100 sampled test images per model with the reference runtime. All matched.