JetBrains has made its AI model for code generation, Mellum, publicly available on the Hugging Face platform and open-sourced the code.
Mellum is designed for code autocompletion, trained on ~4.2 trillion tokens across multiple programming languages with 4 billion parameters and an 8192-token context window.
The training data for Mellum included code snippets from GitHub and articles from Wikipedia, and the training process took approximately 20 days on a cluster of 256 H200 Nvidia GPUs.
JetBrains highlights that Mellum can be integrated into professional development tools, used for AI-based coding assistance, and in educational applications, with the model protected under the Apache 2.0 license for creating derivative products.