No browser target
There is no WASM build. If you need inference inside a web page, this is the wrong tool and no amount of configuration will change that.
For developers
Build on QVAC.
The SDK is open source under Apache 2.0 and installs from npm. The quickstart takes about five minutes. This page collects the things worth knowing before you start, and the ones nobody tells you until you hit them.
Ten minutes in
Install the SDK and run a completion. Everything else is a variation on this.
npm i @qvac/sdkThe API is three verbs: load a model, do something with it, unload it. The same code runs on Node.js, Bare, and Expo.
import { loadModel, completion, unloadModel } from "@qvac/sdk";
const modelId = await loadModel({
modelSrc: "...",
modelType: "llm",
});
const result = await completion({
modelId,
messages: [{ role: "user", content: "Hello" }],
});
await unloadModel({ modelId });If you already have something that speaks to OpenAI's API, you do not need to touch the SDK at all. The QVAC CLI ships an OpenAI-compatible HTTP server on port 11434 — Ollama's default port — so pointing an existing client at local inference is a one-line change to a base URL.
Things that will catch you
Physical devices only
Mobile development does not work on emulators. You will need real hardware from the start, on both platforms.
Bare does not auto-register
Node and Expo spawn a worker that registers every built-in plugin. Bare
runs in-process and registers nothing. Register before your first call
or you will see WorkerPluginsNotRegisteredError.
Bundle size is your problem
Omit the plugins array and the SDK bundles all of them. Name
the ones you use. On mobile this is the difference between a shippable
app and a rejected one.
Intel Macs are CPU-only
No GPU acceleration on macOS x64. Apple silicon or nothing, if you care about speed.
No Python client
It is on a roadmap rather than in a release. Today the SDK is JavaScript and TypeScript. A Python ML researcher gets nothing from it yet.
What Tether is paying for
tether.dev →Tether runs a task-based grants program with no total cap. Each grant is a defined deliverable with a fixed payout and a deadline, paid in USD₮ or Bitcoin. Press coverage put typical individual payouts between $1,500 and $4,000, though the QVAC-specific bounties currently listed run higher.
- LTX-2 Video Generation — qvac-ext-stable-diffusion.cpp + Bare Addon
- ANE Acceleration for llama.cpp — ggml CoreML Backend + QVAC Integration
- QVAC SDK — Swift Client
Observed on tether.dev and reproduced here as a pointer, not a source of truth. Amounts, deadlines, and availability change. Check the source before you plan work around any of it.
Resources
- Documentation ↗ The SDK reference, capability guides, and tutorials for Electron and Expo. Start at the quickstart.
- Source ↗ The monorepo. Apache 2.0. Packages, addons, and the bare-sdk distribution live here.
- Models ↗ MedPsy, the Genesis datasets, and the rest of Tether AI Research's open releases.
- Discord ↗ Where the developers are. The fastest route to an answer about the SDK.
- Grants and bounties ↗ Task-based funding. Pick a deliverable, complete it, get paid.
Get listed here
If you build something on QVAC, it belongs in this directory. Every entry is a markdown file; adding one is a pull request against the repository, and there is a template with the fields filled in.
There is no featured slot, no sponsorship, and no way to buy placement. Released apps sort above unreleased ones and alphabetically within each group, which is enforced in code rather than by hand. The directory is maintained by localhost/AI, who also builds apps listed in it — see About for what that means and does not mean.
Corrections from developers about their own software are taken at their word.