Calcine

Your Snapdragon NPU, as a local AI server.

A desktop app for Qualcomm GenieX: download language models, chat with them on the Hexagon NPU, and give every app on your PC a secure, OpenAI-compatible API.

Website · Download · Security model · Architecture

Why Calcine

GenieX runs language models on the Snapdragon NPU, GPU and CPU from the command line, and geniex serve exposes them over HTTP with no authentication and any web origin allowed. Calcine puts every GenieX command in a friendly window and puts a gateway in front of the server, so you can hand your models to other apps without handing them to every web page you visit.

• Chat with every option. Thinking, temperature, max tokens, compute unit and power mode, adapted to each model's runtime, with tokens per second. Images and voice for vision models.
• Find models that fit. Qualcomm AI Hub models compiled for your chipset, and GGUF models from Hugging Face, with precisions, sizes, memory and disk checks. Paste any Hugging Face, ModelScope or Docker Hub link, or import a folder.
• See the NPU work. Live NPU, GPU, CPU and memory load, drivers, and a self-test that compares speed on each compute unit.
• OpenAI-compatible API. On 127.0.0.1:18181, GenieX's default port, with per-app keys, a request queue and a request log that never records prompts.
• Always up to date. GenieX ships in the installer and can be updated or rolled back from the app. Calcine updates itself with signed packages.
• English and French, light and dark themes, a command palette (Ctrl K) and a tray icon to keep the API running in the background.

Discover: the right precision for your memory	Hardware: live load on every compute unit
Library: your models, ready to run	Server: the API, keys and request log

Get started

1. Download the installer from Releases (Windows 11 on Snapdragon, ARM64). It also installs GenieX.
2. Open Calcine and download a model: the welcome screen suggests a few that fit your PC.
3. Chat with it, or create an API key in Server › API keys and connect an app.

Note

Calcine is in beta and its installers aren't code-signed yet, so Windows SmartScreen asks for confirmation before installing (More info › Run anyway). Releases list SHA-256 checksums, and stable releases also carry an SBOM and a build provenance attestation (gh attestation verify <file> -R eraflo/Calcine).

Use the API

Any OpenAI client works. Point it at Calcine and pass a key created in the app:

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:18181/v1", api_key="calcine_…")
stream = client.chat.completions.create(
    model="qualcomm/Qwen3-4B",
    messages=[{"role": "user", "content": "Hello from my NPU!"}],
    stream=True,
)
for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

curl http://127.0.0.1:18181/v1/chat/completions \
  -H "Authorization: Bearer $CALCINE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qualcomm/Qwen3-4B", "messages": [{"role": "user", "content": "Hello!"}]}'

Clients already written for geniex serve keep working on the same port. For those that can't send a key, Require an API key can be turned off: keyless callers can then only run models.

How it works

flowchart LR
    apps["Your apps<br/>OpenAI SDKs, curl, local web UIs"] -->|"HTTP + API key"| gateway
    ui["Calcine window<br/>chat, library, hardware"] --> gateway
    subgraph calcine ["Calcine"]
        gateway["Gateway · 127.0.0.1:18181<br/>Host, Origin, key and body checks"]
    end
    gateway -->|"loopback, random port"| serve["geniex serve"]
    serve --> npu["Hexagon NPU<br/>QAIRT"]
    serve --> gpu["Adreno GPU · Oryon CPU<br/>llama.cpp"]

Loading

The gateway listens on loopback only, rejects DNS rebinding and browser origins you haven't allowed, checks per-app keys (stored as SHA-256 fingerprints, scoped to inference or management), and refuses local file paths and internal URLs in request bodies. GenieX itself is never exposed, and a Windows Job Object stops it with Calcine, even after a crash. Details in docs/security-model.md.

Requirements

• A Snapdragon PC running Windows 11 on Arm (Linux on ARM64 is planned)
• Free memory and disk space for the models you use: Calcine shows what fits before you download

GenieX is installed by the Calcine installer; nothing else is needed.

Development

Tauri v2 · Rust · React + TypeScript + Vite · Tailwind CSS · bun.

bun install
bun run app:mock   # runs anywhere, with sample data and no GenieX
bun run app        # on a Snapdragon device with GenieX installed

See CONTRIBUTING.md, docs/architecture.md and docs/RELEASING.md. The website lives in site/ and is published to GitHub Pages from main.

License

MIT. Calcine redistributes GenieX under its BSD 3-Clause license, see THIRD_PARTY_NOTICES.md. Calcine is an independent project, not affiliated with Qualcomm.

_{Screenshots show Calcine's demo mode (bun run app:mock), with sample data shaped like a Snapdragon X Elite laptop.}