JetBrains, renowned for its developer-centric tools like IntelliJ IDEA and PyCharm, has introduced Mellum, a newly open-source AI model engineered for code generation and completion. Mellum is now publicly accessible on Hugging Face under the Apache 2.0 license, giving developers and researchers full freedom to experiment, contribute, and integrate the model into their workflows.
A Closer Look at Mellum
Mellum is a specialized large language model (LLM) with 4 billion parameters, trained to excel at code-centric tasks. Originally built to support JetBrains’ internal AI Assistant, Mellum has now been released to the open-source community to spur innovation in software development and machine learning.
What Sets Mellum Apart
- Code-Specific Design: Unlike general-purpose LLMs, Mellum was crafted with software development in mind. Its architecture and training make it more precise and reliable for coding tasks.
- Language Versatility: It can autocomplete and understand multiple programming languages, including Python, Java, Kotlin, Go, and PHP.
- Balanced Performance: With an 8K token context window, Mellum provides a solid mix of output quality and resource efficiency, suitable for both local machines and cloud platforms.
Behind the Model
The model’s training data includes a massive corpus of over 4 trillion tokens drawn from permissively licensed code repositories and English-language content. Built on a LLaMA-style architecture, JetBrains trained Mellum trained using 256 Nvidia H200 GPUs over a 20-day period, utilizing advanced techniques like bf16 precision and Automatic Mixed Precision.
Use Cases and Integration
Mellum is flexible and easy to deploy across environments. It supports inference engines like vLLM for cloud, and llama.cpp or Ollama for offline or embedded setups. Potential uses include:
- Boosting IDEs: Integrating Mellum into IDEs for real-time, context-aware code suggestions.
- Developing Coding Assistants: Powering intelligent tools that help write, review, or refactor code.
- Educational Platforms: Enabling students and instructors to explore AI-supported learning environments.
- Research Projects: Serving as a foundation for experiments in AI-driven programming and LLM fine-tuning.
Embracing Open Innovation
By releasing Mellum to the public, JetBrains is signaling a strong commitment to open research and collaborative progress. The move offers developers and AI practitioners a robust, purpose-built model they can study, adapt, and expand upon.
