Important

I don't use this anymore since 2024.2. This gist should however still be fully valid, except for the missing new data.
Supported languages seem to have changed since 2024.3, but the instructions should still be mostly working.
I'll continue updating the instructions based on the comments on this gist if they change, but you should refer to them for the actual data provided by people. Don't hesitate to comment too to contribute!

Introduction

IntelliJ 2024.1 introduced Full Line Code Completion, available as a bundled plugin in your IDE.

Available at Settings > Editor > General > Inline Completion, you can choose to enable them and download them.

For 2024.1, the Settings path is Settings > Editor > General > Code Completion under Machine Learning-Assisted Completion

Offline installation

If you work on an offline machine, there is no official way to install those models. But after digging into the source code and inspecting the logs with @Kilehynn, we managed to find a way.

Available models

By default, Java and Kotlin models are already installed. "CSS-like" and "JavaScript/Typescript" are the same model internally (it's called "WS", for WebStorm), and "Go" is separate.

Python, PHP and Ruby are also available, but we don't use them so we didn't search the URLs and UUIDs for them.

Installation of new models

To begin, close IntelliJ and head over to your IntelliJ system directory.

1. Move into the full-line directory, then into models. Here, you'll see two UUIDs, one for Java and the other for Kotlin models. models.xml is used by the IDE as a list of installed models.
2. Each model is structured in the following way:
   1. Their folder is an UUID (like the existing models), common for every installation. See info.md to figure out the one for your model.
   2. They contain 3 flcc.* files (flcc.bpe, flcc.json, flcc.model), which AFAIK tell the model how to work
   3. They contain a ready.flag file, an empty file probably indicating the IDE the model is successfully downloaded
   4. Finally, a full-line-inference.zip_extracted directory, containing the executable model as well as its signature in a sign file.
3. For each model, download the flcc.* files and the model.xml if it exists by downloading https://download-cdn.jetbrains.com/resources/ml/full-line/models/[Name]/[Version]-native-llama-bundle/local-model-[Name]-[Version]-native-llama-bundle.jar - and extracting the contents of the [Name]-[Version]-native-llama-bundle folder (new in 2024.2) - like you would with a .zip file. (Replace the values in the URL by the one in the info.md table.)
4. For each model, create an empty ready.flag file
5. Finally, get the models by copying the full-line-inference.zip_extracted folder from the Java or Kotlin model into each new model. Yes, all models are the same. If you want to download them manually yourself, you can use the URL: https://download-cdn.jetbrains.com/resources/ml/full-line/servers/[FLI_version]/[ARCH]/full-line-inference.zip (2024.1 URL required the OS: https://download-cdn.jetbrains.com/resources/ml/full-line/servers/[FLI_version]/[OS]/[ARCH]/full-line-inference.zip ; tested OS and arch in the OS_arch.md file)
6. Finally, you have to add your new model(s) to the models.xml file as such:

<models version="1">
  ...
  <model>
    <version>[Version]-native-llama-bundle</version>
    <size>[Size]</size>
    <languages>
      <language>[One of Languages per line]</language>
      ...
    </languages>
    <binary>flcc.model</binary>
    <bpe>flcc.bpe</bpe>
    <config>flcc.json</config>
    <native>
      <archive>full-line-inference.zip</archive>
      <version>1.0.84</version> <!-- see second table in `info.md` -->
    </native>
    <changelog>[Some text]</changelog>
  </model>
  ...
</models>

An example is available in the attached models_windows_example.xml (IntelliJ doesn't seem to care about size or changelog, you can probably leave them empty/use the same for every installation)

Finally, restart IntelliJ and check the settings, installed languages should all be checked!

Disclaimer: despite all our efforts, some info might be missing or inaccurate, the overall idea should be working though