ModelJars

Model marker JARs for JVM AI applications.

Pin model metadata with Maven-style dependencies, discover descriptors from the classpath, and keep large model weights outside application artifacts. 

dependencies {
    runtimeOnly(
        "org.modeljars.huggingface:" +
        "ggml-org.qwen3-0.6b-gguf.q4_0:" +
        "3.0.0-q4_0.1"
    )
}

Catalog

Search marker coordinates, sources, architectures, quantizations, and backends.

0 models

Runtime

Resolve from the classpath

ModelJars descriptors live in META-INF/modeljars/registry.properties. JVM libraries can load every marker JAR already on the application classpath. 

ModelJarRegistry registry =
    ModelJarRegistry.fromClasspath();

Descriptor

Weights stay external

Marker JARs carry source URLs, checksums, licenses, local cache paths, capabilities, and backend compatibility. They do not embed multi-gigabyte model weights.

Publishing

Catalog changes are reviewed

Catalog metadata is reviewed through GitHub pull requests. CI validates descriptors, generates marker artifacts, and publishes immutable versions once release automation is enabled.

Gradle

Pin a model marker

runtimeOnly(
    "org.modeljars.huggingface:" +
    "ggml-org.qwen3-0.6b-gguf.q4_0:" +
    "3.0.0-q4_0.1"
)

Maven

Use the same coordinates

<dependency>
  <groupId>org.modeljars.huggingface</groupId>
  <artifactId>ggml-org.qwen3-0.6b-gguf.q4_0</artifactId>
  <version>3.0.0-q4_0.1</version>
</dependency>

Contribute

Add a model descriptor

  1. Open a PR against the catalog metadata.
  2. Include upstream source, license, format, quantization, and backend claims.
  3. Let CI validate the generated marker artifact.