Hanzo Engine

Canonical inference + embedding engine for the Hanzo stack, built on top of mistralrs-core.

Consumers (hanzo-vm precompiles, RPC handlers, agents) call [infer] / [embed] which dispatch through a process-wide registry. At startup the runtime registers one [InferenceEngine] and one [EmbeddingEngine] — typically a [MistralEngine] loaded from a Hugging Face repo or a local path.

use std::sync::Arc;
use hanzo_engine::{MistralEngine, register_inference_engine, infer};

# async fn boot() -> anyhow::Result<()> {
// 1) Load a model
let engine = MistralEngine::from_hf_repo("Qwen/Qwen3-4B").await?;
let id = *engine.model_id();

// 2) Register it globally
register_inference_engine(Arc::new(engine))?;

// 3) Sync call sites (EVM precompiles, etc.) dispatch through `infer`
let bytes = infer(&id, b"What is Rust?").unwrap();
println!("{}", String::from_utf8_lossy(&bytes));
# Ok(())
# }

Why a registry?

The hanzo-vm precompiles 0x0201 (AI inference) and 0x0202 (AI embedding) run in a synchronous EVM context. They cannot pass a model handle through the EVM stack. Instead they look up the registered engine at call time. The 32-byte model_id argument lets the engine validate that the requested model is the one it has loaded.