Think

AI & Inference

Chat with an LLM, run classical ML inference synchronously, and compute over encrypted data — all as native precompiles a contract can call directly.

◇Chat replies are simulated locally — no LLM was called. The classifier below is a real, separately trained ONNX model running via WebAssembly, standing in for the on-chain Classical Inference precompile.

LLM Chat

Interactive simulation0x0802· zai-org/GLM-4.7-FP8

Replies are canned per category (Demo Mode), but which category is picked is a real ONNX intent classifier running locally — not keyword matching.

Ask me anything — responses stream in token-by-token, simulating the LLM precompile.

Classical Inference (ONNX)

Real local computation0x0800· synchronous

Synchronous — executes inline in the same block, no TEE round trip. This demo runs a real, trained softmax-regression sentiment model (141-word vocabulary) as genuine ONNX inference, entirely in your browser via WebAssembly — not a hardcoded wordlist.

FHE Inference

Concept / reference0x0807· reference only

Runs ML inference directly on CKKS-encrypted tensors inside a TEE — the executor never sees the plaintext input or output. Long-running (two-phase): submit encrypted input, get an encrypted result via callback. Useful when even the TEE operator shouldn't see your data. No interactive demo here — it's a specialized primitive most dApps reach for only when handling sensitive numerical data (health, finance) that must stay encrypted end-to-end.