JetBrains has open-sourced Mellum, its AI code generation model, available on Hugging Face
JetBrains has launched Mellum, its first open AI code generation model, on the Hugging Face platform. Designed for code completion tasks, Mellum supports numerous programming languages such as Java, Python, and C++, among others, and it's currently integrated into the JetBrains AI Assistant.
The LLM has 4 billion parameters and was trained on over 4 trillion tokens (around 120 billion lines of code) using 256 Nvidia H200 GPUs over 20 days. Licensed under Apache 2.0, Mellum uses permissively sourced datasets like GitHub code and English Wikipedia, making it suitable for both academic and commercial purposes.
JetBrains highlights Mellum's integration potential into developer tools, coding assistants, research, and education, rather than general AI applications. However, the model requires fine-tuning for specific tasks and is not immediately usable. JetBrains offers sample fine-tuned models for Python for experimentation, but these are not for production use. Users are advised that Mellum may reflect biases from public codebases and does not ensure security or vulnerability-free outputs.
- Large Language Model (LLM) Tool
- Free • Open Source
Related news
External links
- Mellum Goes Open Source: A Purpose-Built LLM for Developers, Now on Hugging FaceThe JetBrains Blog • Official source
- JetBrains releases Mellum, an 'open' AI coding modelTechCrunch
Comments
That's a small to medium sized model, that have nearly zero commercial value to a company as big and respected as JetBrains, that are offering good quality IDEs. But the Apache licensing is still a neat gift to community, certainly useful for some medium hobbyist projects, since it could run on any decent computer and AI-accelerated single boards.