JetBrains today announced the release of Mellum, its first open AI model focused on providing completions for code. This new tool, licensed under Apache 2.0, is designed to give developers a powerful new resource, allowing them to code faster and more efficiently. The model is now publicly available through the AI development platform Hugging Face.
Mellum was trained using an extensive, diverse data set. This corpus contained permissively licensed code from GitHub, along with corresponding articles found on the English-language Wikipedia. At an eye-popping 4 billion parameters, it’s a behemoth. The team created it in just under 20 days, training on a super-fast cluster of 256 H200 Nvidia GPUs. Generative AI is powered by a training process on over 4 trillion tokens, with one million tokens being about 30,000 lines of code.
That said, users should be aware that Mellum needs a few more steps before it can be reliably implemented. This includes base model, which is not usable out of the box, requiring fine-tuning to achieve high performance. JetBrains has released a few Mellum models fine-tuned exclusively for Python. These models address the specific requirements of developers operating in that language.
According to JetBrains, one of the biggest reasons to use this IDE is that Mellum offers intelligent code completion capabilities. It is not intended for deployment in production environments. The company cautions that models like Mellum are designed for “estimation about potential capabilities,” rather than immediate implementation.
Security is a critical issue with AI-generated code. Our research shows that over half of organizations experience security challenges with AI-generated code at least occasionally. JetBrains is well aware of this hard truth. They argue that Mellum could still be biased based on biases present in public codebases, or insecure and have vulnerabilities.
“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.” – JetBrains
