JetBrains has made waves recently with the announcement of Mellum, its first open AI coding model, now accessible on Hugging Face. Designed specifically for code generation, Mellum stands out with its impressive training on over 4 trillion tokens and boasting 4 billion parameters, tailored for efficient code completion within various development suites.
This innovative model was trained for approximately 20 days using a formidable setup of 256 H200 Nvidia GPUs, incorporating a wide array of permissively licensed code from platforms such as GitHub, along with data derived from English-language Wikipedia articles. The end result is a versatile tool intended for integration into professional development environments, providing intelligent suggestions and facilitating research in code understanding and generation.
However, potential users should note that while JetBrains has provided fine-tuned versions of Mellum, particularly for Python, these are ideally suited for testing and educational purposes. The company warns not to deploy these initial setups directly in a production environment, underscoring the importance of further fine-tuning for robust applications.
Mellum not only reflects the significant strides AI technology has made in streamlining software development but also raises new concerns regarding security implications. According to a recent survey conducted by Synk, over half of organizations frequently encounter security issues related to AI-generated code. JetBrains has pointed out that Mellum’s outputs may include biases reflective of its underlying training data, which can potentially lead to insecure coding practices or vulnerabilities in the software it helps generate.
In a recent blog post, JetBrains asserted, “This is just the beginning. We’re not chasing generality — we’re building focus. If Mellum sparks even one meaningful experiment, contribution, or collaboration, we would consider it a win.” As AI continues to evolve, the introduction of Mellum heralds a new chapter in the landscape of coding assistance and automation in software development. For those interested, the technical report and additional details can be found on Hugging Face’s platform and JetBrains’ official communications.