JetBrains has released its first open AI model for code generation called “Mellum.” The model is now available to the community on the Hugging Face platform — this allows developers to integrate it into their own tools and experiment with it in various environments. Mellum was trained on over four trillion tokens and contains four billion parameters, with its main application being code suggestions and completions based on context.
JetBrains reports that Mellum’s development took about twenty days using two hundred and fifty-six Nvidia H200 GPUs. Open datasets were used for training, including code from GitHub with open licenses and articles from the English-language Wikipedia, and the model itself is distributed under the Apache 2.0 license.
Mellum is designed for integration into professional developer tools, creating code suggestion assistants, exploring code understanding and generation, as well as for educational purposes and fine-tuning experiments. Currently, JetBrains has provided several Mellum variants specifically fine-tuned for Python, but emphasizes that these versions are not intended for use in production environments and are rather meant to assess the model’s potential.
The company notes that Mellum may reflect certain biases inherent in public codebases and does not guarantee the security or absence of vulnerabilities in the generated code. JetBrains stresses that the goal is to create a focused tool that can be useful for experimentation and collaboration in the field of generative AI.
