JetBrains, a renowned company known for its widely-used software development tools, has publicly unveiled Mellum, its first openly available AI model specifically created for coding applications.
On Wednesday, JetBrains released Mellum onto Hugging Face, a popular AI development platform. Mellum was initially developed and deployed internally last year as a code-generation solution across various JetBrains software suites. The model has been trained on a massive dataset consisting of more than four trillion tokens—equivalent to roughly 120 billion lines of code—and includes approximately 4 billion parameters. These parameters represent the complexity and problem-solving capabilities of the model, while tokens are segments of data the neural network processes during training.
Mellum is specifically designed to assist developers by offering sophisticated code completion—automatically generating code suggestions based on existing context within software projects. According to JetBrains, the AI model suits a variety of professional applications: integration into IDE (Integrated Development Environment) systems for providing intelligent code suggestions, powering AI-assisted coding tools, serving as research infrastructure for studies on software comprehension and generation, and supporting educational software development or finetuning experiments.
Under the permissive Apache 2.0 license, Mellum was trained on openly licensed source code, primarily sourced from GitHub repositories, along with a selection of English-language articles from Wikipedia. To complete this extensive training operation, JetBrains utilized a compute cluster equipped with 256 Nvidia H200 GPUs. The entire training process required around 20 days to finish.
However, Mellum will require extra configuration effort before deployment, as the baseline model is not immediately operational out-of-the-box. It must be fine-tuned for specific development use cases prior to practical application. JetBrains has released a limited number of Mellum models already fine-tuned to assist with Python code but cautioned that these models are primarily designed for evaluating potential rather than immediate use in live production environments.
While AI-generated code tools have rapidly reshaped software development processes, they have also introduced novel security challenges. A recent 2023 survey found that more than half of organizations encountered security issues in AI-produced code either occasionally or frequently.
Acknowledging these challenges, JetBrains openly notes in its accompanying documentation that Mellum may reflect biases found widely in public codebases—such as stylistic preferences or pre-existing security issues present in open repositories. The company stresses that code suggestions provided by Mellum cannot reliably be considered secure or free of vulnerabilities until thoroughly vetted by human experts.
JetBrains emphasizes that Mellum represents the start of a long-term initiative focused on specialized AI models tailored specifically to practical and professional developer uses, rather than pursuing broad general-purpose AI capabilities. In a corporate blog announcement, the company expressed hope that Mellum would inspire further experimentation and collaboration within the greater software development community.
