Kotlin Multiplatform bindings for lingua-rs 1.8.0 (75 languages). Android and Desktop JVM use JNI; iOS uses Kotlin/Native cinterop and a Rust static library. All use the original Rust models. No Java language-detection implementation is used.

import com.zhangke.lingua.*

val detector = LanguageDetectorBuilder
    .fromAllLanguages()
    .withMinimumRelativeDistance(0.25)
    .build()

try {
    val language = detector.detectLanguageOf("今日はいい天気ですね")
    val confidence = detector.computeLanguageConfidence("Hello world", Language.ENGLISH)
    val values = detector.computeLanguageConfidenceValues("Hello world")
    val sections = detector.detectMultipleLanguagesOf("Hello world. 今日はいい天気ですね")
} finally {
    detector.close()
}

LanguageDetector implements AutoCloseable, so build().use { detector -> … } is also supported. Keep a detector for repeated calls. Operations block; run them off the UI thread on Android, iOS, and desktop. The multiplatform demo provides an executable example for every detector method on Android, iOS, and desktop, using a separate detector for each example.

See the audited API map for the complete upstream mapping. The API includes language and ISO enums, all builder factories/options, detection, confidence, mixed-language segmentation, the four native parallel batch methods, model unloading, and upstream's four file-generation tools.

• Language exposes isoCode6391 / isoCode6393, ISO lookup and language/script sets.
• ISO enums have lowercase code values and case-insensitive fromCode parsing.
• Builder subsets require at least one language; duplicates are removed.
• Relative distance must be finite and in 0.0..0.99 (including endpoints).
• Low accuracy mode uses less memory, with lower accuracy for short texts.
• detectLanguageOf returns null when no language can be reliably detected.
• Confidence values are LanguageConfidence(language, value) in descending order. Equal-confidence ordering is unspecified, as upstream; empty/unclassifiable text can produce all-zero values. A language outside the subset has confidence zero.
• Batch methods preserve input order and invoke upstream's Rayon parallel methods.
• DetectionResult(startIndex, endIndex, wordCount, language) uses UTF-16 indices, with an exclusive end, suitable for text.substring(startIndex, endIndex). Rust's original UTF-8 byte offsets are converted, including supplementary characters.

Android, Desktop, and iOS serialize calls and close() per detector. Parallel batch operations still use native worker threads; separate detector instances may run concurrently. close() is idempotent and every operation after close throws IllegalStateException. Android and Desktop use a phantom-reference daemon; iOS uses a Kotlin/Native cleaner as fallbacks for forgotten close calls; deterministic close/use remains preferable.

The Rust registry stores Arc<Detector> behind opaque, monotonically increasing IDs; IDs are never raw pointers or reused. An in-flight operation retains an Arc if its registry entry is removed. Upstream LanguageDetector: Send + Sync is compile-checked. Its model caches are process-wide LazyLock<DashMap<…>> values. Closing releases the detector, not the shared model caches. unloadLanguageModels() removes shared models for the configured languages; other detectors can reload those models on demand.

All JNI exports catch unwinding Rust panics and translate them to IllegalStateException. Invalid builder arguments, language IDs, null native inputs and unpaired UTF-16 surrogates produce IllegalArgumentException. Invalid/stale native handles produce IllegalStateException. Existing JVM exceptions, including OutOfMemoryError, are preserved. Fallible registry allocation reports OutOfMemoryError. Rust/lingua's infallible allocator may abort on process-wide OOM; catch_unwind cannot recover an allocator abort. No promise of recovery from native OOM is made.

The file writers use absolute filesystem paths and existing output directories, may overwrite their output files, and accept UTF-8 input. Paths are not Android content URIs. UInt preserves upstream's u32 count domain. Argument errors are validated before writing; I/O errors map to IllegalStateException. Unexpected upstream panics also map to IllegalStateException. Generated model files do not replace the detector's embedded models automatically. Generating unique ngrams for all 75 languages is an offline, memory-intensive operation.

• Android: JNI implementation, arm64-v8a and x86_64, minSdk 23.
• iOS Arm64 / iOS Simulator Arm64: C ABI and Kotlin/Native cinterop, all detector and file-writer operations implemented. Builds require full Xcode; the current Kotlin/Native configuration targets iOS 15 or later.
• Desktop JVM: JNI implements all detector and file-writer operations. macOS ARM64 is verified locally. Build/loading supports macOS, Linux (GNU libc), and Windows (MSVC), each with ARM64 and x86_64 native classifiers. A six-job CI matrix builds and tests those classifiers on matching host architectures.

native/src/core.rs, language.rs and writers.rs are platform-independent. jni_bridge.rs handles JNI conversion and exceptions for Android and Desktop. src/jvmSharedMain shares JNI declarations, result decoding, file writers, and serialized ownership across both JVM platforms; platform source sets load the library. c_bridge.rs exposes that same core through the C interface in native/include/lingua.h. iOS builds exclude JNI and use src/iosMain to implement the common Kotlin API.

The C ABI borrows length-delimited UTF-16 strings for each call, preserving NULs and rejecting malformed surrogates. Every operation returns an owned opaque result; Kotlin copies its contents and releases it in finally. Confidence and segment rows follow the existing wire mapping, with integers represented exactly as doubles. Handles remain separate 64-bit integers. Rust errors and unwinding panics cannot cross the C boundary: argument errors become IllegalArgumentException, while state, panic, and fallible allocation errors become IllegalStateException on iOS. An allocator abort remains unrecoverable.

The private wire ID packs the two lowercase ISO 639-1 ASCII bytes as (first << 8) | second; zero means no detected language. It does not depend on Rust/Kotlin enum ordinals. Confidence payloads contain alternating ID/value doubles; segment payloads contain groups of four ints. Nested batches use primitive-array rows. No JSON or public handle types are involved. Consumer ProGuard rules retain the JNI class and exported method names.

tools/languages.json is the reviewed lingua 1.8.0 metadata manifest. Regenerate both language bindings with python3 tools/generate_language_models.py, followed by cargo fmt --manifest-path native/Cargo.toml. Rust tests cross-check every language, ISO code and script group against upstream and the generated Kotlin file.

Prerequisites: JDK 21, Android SDK/NDK, Cargo, cargo-ndk, and the aarch64-linux-android / x86_64-linux-android Rust targets.

cargo test --manifest-path native/Cargo.toml --locked
./gradlew :lingua-core:build
./gradlew :demo:androidApp:connectedDebugAndroidTest

buildRustAndroid keeps the existing cargo ndk → jniLibs → AAR chain and uses --locked. Both Android ABIs are packaged in lingua-core/build/outputs/aar/lingua-core.aar. Generated .so files are ignored by Git; native/Cargo.lock is tracked.

commonTest runs through Desktop. iOS KLIBs still compile on a Command Line Tools host; iOS simulator test linking/execution is skipped when the simulator SDK is absent and enabled when it is installed. Android tests require a connected device or emulator. They cover detection, Latin-script discrimination, confidence, parallel APIs, Unicode offsets, lifecycle/concurrency, JNI validation, and file writers. Full all-language unique-ngram generation is not exercised in routine tests because of its resource cost.

Upstream 1.8.0 limitation: TestDataFilesWriter never terminates when maximumLines is zero or exceeds the input line count. The bridge deliberately rejects those values with IllegalArgumentException before writing. This is a safety correction to an upstream edge case, not an alternative writer algorithm.

Install full Xcode and select its developer directory. Install the Rust targets:

rustup target add aarch64-apple-ios aarch64-apple-ios-sim
./gradlew :lingua-core:compileKotlinIosArm64 :lingua-core:compileKotlinIosSimulatorArm64
./gradlew :lingua-core:iosSimulatorArm64Test
./gradlew :lingua-core:assembleLinguaCoreReleaseXCFramework

Gradle builds the Rust static library separately for each target before cinterop, then embeds it in the generated interop KLIB. KMP consumers therefore link the Rust library automatically; they do not need to install Cargo to consume a published KLIB. For Swift consumers, the last command produces the static lingua-core/build/XCFrameworks/release/LinguaCore.xcframework. Explicitly close detectors from Swift when finished; expensive methods are synchronous. Detector, builder validation, and writer entry points declare @Throws so their expected Kotlin exceptions are exported as errors Swift can catch.

The iOS tests cover all detector methods, parallel ordering, empty batches, confidence values, builder options, UTF-16 offsets including emoji, embedded NULs, invalid surrogates, close semantics, file generation, and native error/panic translation. Native C ABI tests also run with cargo test on the host without an iOS SDK.

import LinguaCore

let detector = try LanguageDetectorBuilder.companion.fromAllLanguages().build()
defer { detector.close() }
let language = try detector.detectLanguageOf(text: "今日はいい天気ですね")

Development requires JDK 21, Cargo, the selected Rust target (rustup target add <target>), and the native platform linker: Xcode Command Line Tools on macOS, GCC on Linux, or Visual Studio C++ Build Tools with the selected architecture on Windows. The generated JVM classes target Java 11. Gradle compiles a Rust dynamic library for the current host and stages it as natives/<os>-<arch>/<library>. The loader extracts this classpath resource to a unique temporary directory and loads it once per class loader. Explicit close/use is recommended; a phantom-reference reaper provides a GC cleanup fallback.

./gradlew :lingua-core:desktopTest :lingua-core:desktopPackagedTest
./gradlew :lingua-core:desktopJar :lingua-core:desktopNativesJar

For macOS ARM64, these produce two artifacts under lingua-core/build/libs/:

• lingua-core-desktop-0.0.1.jar: portable JVM classes, without native binaries.
• lingua-core-desktop-0.0.1-natives-macos-aarch64.jar: the host Rust/JNI library.

Both artifacts are attached to the Desktop Maven publication. After publishing, JVM consumers include the classes and their platform's native classifier:

dependencies {
    implementation("io.github.0xzhangke:lingua-core-desktop:0.0.1")
    runtimeOnly("io.github.0xzhangke:lingua-core-desktop:0.0.1:natives-macos-aarch64")
}

KMP consumers can keep the common dependency and add that native classifier to desktopMain. Runtime consumers need neither Cargo nor java.library.path setup. Other recognized classifiers are natives-macos-x86_64, natives-linux-aarch64, natives-linux-x86_64, natives-windows-aarch64, and natives-windows-x86_64. By default a build generates the host JVM architecture's classifier. Select another architecture on the same OS with -Plingua.desktopArch=x86_64 or -Plingua.desktopArch=aarch64; arm64, amd64, and x64 aliases are accepted. Install that Rust target and provide a cross linker when needed. Selecting an architecture does not change the operating system or install a linker.

For example, build the Intel Mac native JAR on an Apple Silicon Mac:

rustup target add x86_64-apple-darwin
./gradlew -Plingua.desktopOnly=true -Plingua.desktopArch=x86_64 \
  :lingua-core:desktopNativesJar

lingua.desktopOnly=true excludes mobile targets and the Android demo so desktop builds do not require Android/iOS SDKs or publication signing credentials. Use the default configuration for publishing the full KMP library; signing is still required for publication and signing tasks. Each invocation attaches its selected classifier to the Desktop publication. CI artifacts are build outputs, not Maven releases; a release must collect the six classifiers under the same Maven version. Place the six native JARs in one directory, preserving their filenames, then set -Plingua.desktopNativeArtifactsDir=/path/to/collected-jars for publication. Gradle requires all six and attaches them to the Desktop Maven publication, replacing the default single-host classifier. Normal signing requirements apply.

Tests must use a JVM matching the native library's architecture. On Apple Silicon, Intel tests require Rosetta and an x86_64 JDK/JRE; pass its home directory:

./gradlew -Plingua.desktopOnly=true -Plingua.desktopArch=x86_64 \
  -Plingua.desktopJavaHome=/path/to/x86_64-jdk/Contents/Home \
  :lingua-core:desktopTest :lingua-core:desktopPackagedTest

The Desktop workflow builds, runs the API/JAR/Rust tests, and uploads each native classifier separately on six matching runners. Runner labels follow GitHub's supported runner matrix. Linux artifacts built on Ubuntu 24.04 require a compatible GNU libc system; Alpine musl, 32-bit x86/ARM, and other architectures are not covered.

desktopTest covers detection, all parallel operations, confidence, UTF-16 offsets, builder options, close/concurrency semantics, file generation, and native errors. desktopPackagedTest reruns the API tests using only packaged production JARs, checks that the native resource comes from a JAR, and uses a separate JVM to verify the missing-native-artifact error. Both macOS ARM64 and x86_64 (using an Intel JVM under Rosetta) have been exercised locally. Linux/Windows execution awaits the CI matrix running on those hosts.

demo uses Compose Multiplatform. The English test UI, inputs, results, theme, background execution, and examples for all ten LanguageDetector methods live in demo/src/commonMain. The entry points only host the shared App():

• Android: demo/androidApp/src/main/kotlin/.../MainActivity.kt.
• iOS: demo/src/iosMain/.../MainViewController.kt, hosted by the SwiftUI app in demo/iosApp/LinguaDemo.xcodeproj.
• Desktop: demo/src/desktopMain/.../main.kt.

AGP 9 requires a separate Android application module, so :demo:androidApp contains only the Android launcher, packaging/resources, and existing Android instrumentation tests. demo supplies its Android library target as well as the iOS ARM64, iOS Simulator ARM64, and desktop JVM targets.

The demo consumes io.github.0xzhangke:lingua-core:0.0.1 from Maven Central, with the version shared through linguaCore in gradle/libs.versions.toml. Desktop adds the published lingua-core-desktop native classifier matching the Gradle JVM's operating system and CPU architecture. Building or running the demo does not build the local lingua-core module or require Cargo/Rust targets; the published Android AAR and iOS KLIBs already include their native libraries. The selected desktop native classifier must exist in the published version.

# Android APK / installation / existing instrumentation tests
./gradlew :demo:androidApp:assembleDebug
./gradlew :demo:androidApp:installDebug
./gradlew :demo:androidApp:connectedDebugAndroidTest

# Desktop window (host native JAR is included automatically)
./gradlew :demo:run

# Execute all ten shared examples and formatting checks on Desktop
./gradlew :demo:desktopTest

# iOS framework compilation with full Xcode selected
./gradlew :demo:linkDebugFrameworkIosSimulatorArm64

For iOS, open demo/iosApp/LinguaDemo.xcodeproj in Xcode and run the LinguaDemo scheme on an ARM64 simulator or device. Its build phase runs :demo:embedAndSignAppleFrameworkForXcode to build/link the shared UI and published native library automatically. An iOS Simulator runtime is needed to launch on a simulator; set a signing team in Xcode to run on a physical device. Xcode's Gradle build phase needs JDK 21 available (JAVA_HOME can be set in your local Xcode environment). No CocoaPods setup is needed.