JetBrains Launches Open-Source Code Model Mellum-4b-base

JetBrains has open-sourced its first large language model specifically designed for code optimization, Mellum-4b-base. This model features 4 billion parameters and is trained on a dataset of 4.2 trillion tokens, with a focus on code completion and IDE integration. Mellum-4b-base was trained on a cluster of 256 NVIDIA H200 GPUs, delivering powerful capabilities in code understanding and generation. In addition to the base model, JetBrains has also released a version fine-tuned for Python. While the model demonstrates strong performance in code completion for Java and Python, its support for other programming languages remains relatively limited.