---
title: "Run it here | Xybrid"
description: "Run SmolLM2 in your browser with Xybrid. Your prompt stays on your device."
url: "https://www.xybrid.ai/demo"
---

# The model runs here.

Your prompt. Your device. Try a small language model on your GPU or CPU, right in this tab.

This demo runs only in a desktop browser tab: the model downloads and generates in the visitor’s browser when they press Run. Fetching this page doesn’t run it.

## SmolLM2 135M

Runs on your device. 136 MB model download on first run. Cached by your browser. Up to 256 output tokens.

Run with:

- Auto — Selects WebGPU when available, with a CPU fallback.
- WebGPU — Uses your GPU. Availability depends on your browser and device.
- CPU (WASM) — Runs on the CPU in a worker, keeping this page responsive.

## Suggested prompts

- Offline haiku — Write a haiku about AI that works offline.
- Why on-device? — Explain in two sentences why on-device inference matters.
- Local advantages — Give three reasons to run a language model locally.

No account. No API key. Inference happens here.

## Build this into your app

`demo.ts`, as the page shows it for the prompt “Offline haiku” and the Auto engine:

```ts
import { XybridLlm } from '@xybrid/web';

const llm = await XybridLlm.load(
  'https://huggingface.co/xybrid-ai/SmolLM2-135M-Instruct-LiteRTLM/resolve/main/model_metadata.json',
  {
    wasmPath: '/xybrid/runtime', // the runtime files, served from your origin
    accelerator: 'auto',
  },
);

const stream = llm.generateStream(
  "Write a haiku about AI that works offline.",
  { maxOutputTokens: 256 },
);
for await (const delta of stream) {
  // Append each delta to your response.
  console.log(delta);
}
```

[Explore the web SDK](https://github.com/xybrid-ai/xybrid/tree/master/bindings/web)

From this tab to your app. [Read the quickstart](https://docs.xybrid.dev/en/docs/quickstart)
