JetBrains released Mellum, its AI-powered code completion model, publicly. Mellum is now accessible on the AI platform Hugging Face under the Apache 2.0 license.
This model aims to assist developers by offering faster, context-aware code suggestions. Mellum was trained using 256 Nvidia H200 GPUs over about 20 days on a dataset of over 4 trillion tokens.
Mellum’s Development and Capabilities
Mellum, designed by JetBrains, targets intelligent code completion within integrated development environments (IDEs). The model has 4 billion parameters, indicating its problem-solving capacity.
The training utilized permissively licensed GitHub code and English Wikipedia articles, enabling Mellum to understand coding patterns effectively. JetBrains offers fine-tuned versions, especially for Python, to demonstrate its capabilities.
Developers need to fine-tune Mellum before use, as the base model is not production-ready. The company notes that generated code may contain biases and security vulnerabilities, urging careful review and testing.
Significance and Usage
Making Mellum open-source encourages community collaboration and innovation in AI-assisted software development. It fits professional tooling, coding assistants, educational use, and research in code generation.
JetBrains highlights Mellum’s focus on coding rather than broad language tasks, stating, “If Mellum sparks even one meaningful experiment, contribution, or collaboration, we would consider it a win” in a blog post.
Key Model Statistics
Feature | Specification |
---|---|
Model Size | 4 billion parameters |
Training Data | Over 4 trillion tokens |
Training Duration | Approximately 20 days |
Hardware | 256 Nvidia H200 GPUs |
Mellum is hosted on Hugging Face, a platform well-known for sharing AI models. This hosting allows developers to access and fine-tune Mellum for their projects.
Security Considerations
AI-generated code introduces security concerns. A recent survey by Synk found over 50% of organizations reported security issues with AI-produced code. JetBrains advises caution when using Mellum’s suggestions.
Leave a Reply