@agentix-e/timesfm-web

Browser inference engine for TimesFM — zero-shot forecasting via WebAssembly, WebGPU, or WebGL.

Overview

@agentix-e/timesfm-web enables TimesFM inference directly in the browser using onnxruntime-web. It supports multiple backends (WASM, WebGPU, WebGL) and loads the 885 MB ONNX model via fetch() with streaming.

Backend Support

| Backend | Browser Support | Performance | | ------- | ---------------------- | ----------- | | WebGPU | Chrome 113+, Edge 113+ | 🚀 Fastest | | WASM | All modern browsers | ✅ Default | | WebGL | Fallback | ⚠️ Slow |

Memory note: The 885 MB model must fit in browser WASM memory (~4 GB limit).

Installation

npm install @agentix-e/timesfm-web onnxruntime-web

onnxruntime-web is a peer dependency — install it alongside.

Quick Start

import { TimesFMModel, createForecastConfig, TIMESFM_25_CONFIG } from '@agentix-e/timesfm-core';
import { TimesFMWebInferenceEngine, loadModelFromUrl } from '@agentix-e/timesfm-web';

// 1. Download the model via fetch() with progress tracking
const { buffer } = await loadModelFromUrl('https://cdn.example.com/timesfm-2.5.onnx', {
  onProgress: (received, total) => console.log(`${received}/${total}`),
});

// 2. Create web engine and load model
const engine = new TimesFMWebInferenceEngine(TIMESFM_25_CONFIG);
await engine.load(buffer);

// 3. Create model with injected web engine (DI pattern)
const model = await TimesFMModel.fromPretrained({
  modelPath: '/models/timesfm-2.5.onnx',
  engine,
});

// 4. Compile and forecast
model.compile(createForecastConfig({ maxContext: 1024, maxHorizon: 256 }));
const result = await model.forecast(24, [
  new Float32Array([
    /* historical values */
  ]),
]);

console.log(result.pointForecast); // Shape: [1, 24]
console.log(result.quantileForecast); // Shape: [1, 24, 10]

await model.dispose();

API Documentation

📚 Full API reference: agentix-e.github.io/timesfm-ts/api/modules/timesfm-web.html

Key exports:

• TimesFMWebInferenceEngine — Browser inference engine implementing IInferenceEngine (WebGPU / WASM / WebGL)
• loadModelFromUrl — Fetch-based model loader with streaming progress callback
• checkModelAvailability — HEAD request check for model availability
• WebEngineLogger — Custom logger interface for browser engine diagnostics

License

Apache 2.0