JetBrains Launches Mellum: Open-Source LLM Revolutionizes AI-Powered Code Completion

JetBrains, a leader in software development tools, has unveiled Mellum-4b-base, its first open-source large language model (LLM) optimized specifically for code-related tasks. Built on the LLaMA architecture and boasting 4 billion parameters, Mellum is trained on over 4 trillion tokens, with an 8192-token context window supporting code completion across multiple programming languages. JetBrains designed Mellum for languages like Python, Java, Kotlin, Go, PHP, C, C++, C#, JavaScript, TypeScript, CSS, HTML, Rust, and Ruby.

Both the base model (Mellum-4b-base) and a Python fine-tuned variant (Mellum-4b-sft-python) are now freely available on Hugging Face under the Apache 2.0 license. With this move, JetBrains aims to give researchers, educators, and advanced engineering teams deep insight into a purpose-built code model, enabling customization and experimentation. While Mellum may not be plug-and-play for the average developer yet, its architecture is ideal for advanced users fine-tuning AI for specialized coding solutions.

Mellum has undergone rigorous benchmarking and human evaluation, performing competitively against other code-centric LLMs. Tests in Python and Java showed stable results across varying code context lengths; notably, its fine-tuned Python version achieved over 80% success in single-line code completion tasks. Mellum’s efficient, focused design delivers expert solutions where general-purpose models fall short.

JetBrains cites “transparency, collaboration, and the power of shared progress” as key reasons for open-sourcing Mellum, hoping to accelerate advancements in code generation and understanding within industry and academia alike. The company also plans to expand the Mellum family with models targeting a broader range of coding tasks, such as diff prediction.