JetBrains Releases Open-Source AI Coding Model, Mellum
JetBrains, creator of popular developer tools, has open-sourced its AI coding model, Mellum, on Hugging Face. Previously integrated into JetBrains' software suites, Mellum is now publicly available.
Trained on over 4 trillion tokens, this 4-billion parameter model specializes in code completion. This means it predicts and suggests code snippets based on surrounding context within a developer's environment.
Mellum's Capabilities and Training
Mellum is designed for integration with professional developer tools, AI coding assistants, and code-related research. It's also suitable for educational purposes and fine-tuning experiments.
Trained on permissively licensed code from GitHub and English-language Wikipedia articles, Mellum's training took approximately 20 days using 256 Nvidia H200 GPUs. The model is available under the Apache 2.0 license.
While the base model requires fine-tuning before use, JetBrains provides pre-trained Python models. However, these are intended for evaluating potential capabilities, not for production environments.
Security Considerations and Future Development
As with other AI-generated code, security is a concern. JetBrains acknowledges that Mellum may reflect biases found in public codebases and its suggestions might not be secure or vulnerability-free.
The company emphasizes that this release is just the beginning. Their focus is on building specialized tools for developers, not general AI solutions. They hope Mellum will inspire experimentation, contributions, and collaboration within the developer community.
For more technical details, see the Mellum model on Hugging Face and the JetBrains blog post.
