JetBrains, the renowned developer behind popular integrated development environments (IDEs) such as IntelliJ IDEA and PyCharm, has made a significant leap in the AI space by open-sourcing its proprietary code-generating model, Mellum. Mellum is now openly available to the global AI and developer community via the Hugging Face platform, marking JetBrains’ first “open” AI model specifically designed for coding tasks.

What Is Mellum? A Focal Model for Code Completion

Unlike general-purpose large language models, Mellum is a “focal model” a term JetBrains uses to describe AI built for deep expertise in a single domain rather than broad, general capabilities. Trained from scratch on more than 4 trillion tokens (equivalent to roughly 120 billion lines of code) and weighing in at 4 billion parameters, Mellum is engineered for code completion: predicting and generating code snippets based on context within professional developer tools.

JetBrains designed Mellum to integrate seamlessly into its suite of IDEs, powering features like intelligent code suggestions, real-time error detection, and context-aware code generation. The model currently supports a wide range of programming languages, including Java, Kotlin, Python, Go, PHP, C, C++, C#, JavaScript, TypeScript, CSS, HTML, Rust, and Ruby.

Why Open Source Mellum?

JetBrains’ decision to open source Mellum stems from a belief in transparency, collaboration, and the transformative power of shared progress. The company draws parallels to the impact of open-source projects like Linux and Git, which have driven major technological advances. By releasing Mellum’s base model under the Apache 2.0 license, JetBrains aims to give researchers, educators, and advanced engineering teams a chance to study, experiment with, and adapt a purpose-built code AI model.

The company emphasizes that Mellum is not a plug-and-play solution for everyday developers. The base model requires fine-tuning before use, and while JetBrains has provided some Python-specialized fine-tuned versions, these are intended for research and estimation, not production deployment.

How Does Mellum Perform?

Benchmark data shared by JetBrains shows Mellum delivering competitive results for its size and specialization, especially in code completion speed and accuracy. While it may not outperform the largest generalist models like CodeLlama-13B in every metric, Mellum’s efficiency and focus make it a strong candidate for integration into developer tools where speed and relevance are paramount.

JetBrains reports that, since Mellum began powering code completion in its AI Assistant, completion latency has dropped to one-third of previous times, acceptance rates of suggestions have reached 40%, and the cancellation rate of suggestions is now three to four times lower than before.

Security and Limitations

JetBrains is transparent about the limitations and risks of AI-generated code. The company acknowledges that Mellum, like all models trained on public codebases, may reflect biases and could generate code that is not always secure or free from vulnerabilities. A 2023 survey by Snyk found that over half of organizations encounter security issues with AI-produced code, highlighting the need for caution and review when integrating AI-generated code into production systems.

JetBrains sees Mellum as just the beginning of a new family of focal models, each specialized for different coding tasks code completion to diff prediction and beyond. The company’s vision is to foster a collaborative ecosystem where researchers and advanced teams can build, experiment, and contribute to the evolution of AI-powered developer tools.