@cohesiumai/providers-wasm

> WASM provider for browser-ai — Universal fallback using WebAssembly.

Installation

``bash
pnpm add @cohesiumai/providers-wasm
`

Usage

`typescript
import { createWASMProvider } from '@cohesiumai/providers-wasm';

const provider = createWASMProvider();

// Check availability (always true in modern browsers)
const result = await provider.detect();
if (result.available) {
console.log('WASM is available');
}
`

Browser Support

| Browser | Support |
|---------|---------|
| Chrome | ✅ Full |
| Edge | ✅ Full |
| Firefox | ✅ Full |
| Safari | ✅ Full |
| Mobile browsers | ✅ Full |

Features

- Universal Compatibility — Works in all modern browsers
- No GPU Required — CPU-based inference
- Reliable Fallback — Always available option
- Smaller Models — Optimized for quantized models

API

$3

Creates a new WASM provider instance.

`typescript
interface WASMProviderOptions {
numThreads?: number; // Worker threads (default: navigator.hardwareConcurrency)
simd?: boolean; // Use SIMD if available (default: true)
}
`

$3

- detect() — Check if WASM is available
- init(config, modelSpec) — Initialize with model
- generate(params) — Run inference
- abort() — Cancel current generation
- teardown() — Cleanup resources
- getDownloadProgress()` — Get model download status

Performance Notes

WASM inference is slower than WebGPU or WebNN but provides:
- Consistent behavior across all browsers
- No GPU memory constraints
- Works on low-end devices

Recommended for:
- Fallback when other providers unavailable
- Small quantized models (Q4, Q8)
- Text-only tasks with moderate latency tolerance

V0.2+

This provider was introduced in browser-ai V0.2.

License

MIT © 2026