Models & Tiers

RustO! separates its core runtime engine from model weights so your production binaries stay under 10 MB and don't carry unnecessary model payload.

If you are unsure which model to pick: start with PP-OCRv6 Tiny (FP32). At ~6 MB, it runs in under 50 ms on mobile CPUs and accurately extracts standard receipts, invoices, and documents. Only deviate to quantized builds or larger models if you hit specific hardware or recall constraints.

1. Which Model Should You Use?

2. Decision Flowchart

3. How to Get the Models

Choose the delivery pattern that matches your application architecture:

Option 1: Direct Zip Releases from GitHub (Universal)

If you are deploying desktop binaries, self-hosting models, or downloading weights on first run, grab the pre-packaged zip archives from GitHub Releases.

Each archive contains the complete triple: det.rten (detector), rec.rten (recognizer), and dict.txt (dictionary).

Release URL: https://github.com/byrizki/rusto-rs/releases/download/models-rten-v0.3.2/

Loading Extracted Files in Code:

Unzip the files into your application's asset directory and point the initializer to them:

// Rust
let config = InitializeConfig::ppv6(
    "assets/models/det.rten",
    "assets/models/rec.rten",
    "assets/models/dict.txt"
);
let mut ocr = RustO::initialize(config)?;

Option 2: Native Package Managers (Zero Manual Downloads)

If you are developing for iOS, Android, or .NET, you do not need to download zip files manually. Add the matching model bundle directly to your project manifest:

iOS (CocoaPods)

In your Podfile:

pod 'RustO-Models-PPOCRv6-Tiny', '~> 0.3.2'

Calling try RustO.initialize() automatically detects and loads the model files inside the pod bundle.

Android (Gradle)

In android/app/build.gradle:

dependencies {
    implementation 'com.github.byrizki.rusto-rs:rusto-models-ppocrv6-tiny:0.3.2'
}

Models are embedded into your APK/AAR assets and extracted at startup.

.NET (NuGet)

dotnet add package RustODotnet.Models.PPOCRv6.Tiny

MSBuild copies the model files straight to your build output directory.

Option 3: CLI Downloader Script (Development & CI)

For local development or CI build pipelines, download and convert models automatically using the provided script:

# Download default PP-OCRv6 Tiny FP32 models
bash scripts/download_models.sh --model ppocrv6 --tier tiny --output-dir models/PPOCR_v6_tiny

# Download quantized Tiny INT8 models (~2.8 MB)
bash scripts/download_models.sh --model ppocrv6 --tier tiny --int8 --output-dir models/PPOCR_v6_tiny_int8

# Download PP-OCRv6 Small models
bash scripts/download_models.sh --model ppocrv6 --tier small --output-dir models/PPOCR_v6_small

# Download Japanese language pack
bash scripts/download_models.sh --model ppocrv4 --tier mobile --lang japan --output-dir models/PPOCR_v4_japan

Option 4: On-Demand Dynamic Download

If your mobile app needs to stay under a strict cellular download cap in the app store, you can ship your app without models and fetch the zip archive on first launch:

import fs from 'fs';
import { initialize } from '@rustors/node';

async function ensureModelsDownloaded(targetDir: string) {
  if (fs.existsSync(`${targetDir}/det.rten`)) return;

  const zipUrl = 'https://github.com/byrizki/rusto-rs/releases/download/models-rten-v0.3.2/RustO-Models-PPOCRv6-Tiny.zip';
  const response = await fetch(zipUrl);
  const buffer = await response.arrayBuffer();

  // Extract buffer into targetDir and initialize engine...
}

4. Format Overview: .rten vs .onnx vs _int8

• .rten (Recommended):
  • Binary flatbuffer format designed for the pure-Rust RTen engine.
  • Zero-copy memory mapping (mmap) allows the operating system to share model pages between processes without extra heap allocations.
  • Pure Rust with SIMD vectorization (AVX2, AVX-512, and ARM NEON) and zero external C++ runtime dependencies.
• .onnx:
  • Useful when running with ONNX Runtime (ort) or targeting WebAssembly in browsers.
• _int8.rten (Quantized):
  • 8-bit integer weights.
  • Reduces file size and memory footprint by ~50–60%.
  • Up to 2x faster memory throughput on mobile memory buses with less than 0.5% character accuracy variance.

5. Checksum Verification

Every release includes a SHA256SUMS file. Verify your downloaded archive with:

sha256sum -c SHA256SUMS --ignore-missing