@ariacompute/engine-ts

npm i @ariacompute/engine-ts

libaria_ffi is resolved as ARIA_FFI_LIB → package-bundled dist/lib/ → ~/.ariacompute/lib/ (same as aria-engine upgrade). If none exist, Engine.open downloads the latest stable libaria_ffi_{ver}_{os}.tar.gz from regional Releases. ARIA_FFI_LIB remains an optional override.

import { Engine } from "@ariacompute/engine-ts";
const eng = new Engine("/path/to/bundle");
console.log(eng.complete([{ role: "user", content: "hi" }]));
eng.close();

Auto-download by model name

Engine.open(modelRef, { token, site }) accepts a local bundle path or an Aria model name (e.g. gemma-4-e2b-it_q4). A value containing / or already on disk is loaded directly; otherwise the SDK downloads it from the regional public hub (same as aria-engine download: .com → Hugging Face, .cn → ModelScope; site defaults to https://ariacompute.com) into ~/.ariacompute/models/{model} and then loads it. Dashboard is not used. A Dashboard sk- / bfvk- token is ignored for hub auth. Token is optional for public models. A valid cached bundle is reused without re-downloading.

Gated files: call eng.setup({ hf_token }) (.com) or eng.setup({ modelscope_api_token, site_url }) (.cn) before open. Same keys as aria-engine setup. Instance-only; does not write engine.yml. If omitted, the SDK reads ~/.ariacompute/engine.yml.

const eng = await Engine.open("gemma-4-e2b-it_q4");

const gated = new Engine();
gated.setup({ hf_token: "hf_..." });
await gated.open("gemma-4-e2b-it_q4");