JetBrains releases Mellum, an ‘open’ AI coding model

JetBrains, the company behind a series of popular app development tools, has released its first “open” AI model for coding.

On Wednesday, JetBrains Mellum, a code-generating model that the company released last year for its various software development suites, made available openly on the AI ​​DEV platform. Mellum, trained on more than 4 trillion tokens, weighs 4 billion parameters and is specially designed for code voltage (ie the completion of code features based on the surrounding context).

Parameters are roughly in line with the problem -solving skills of a model, while tokens are the raw pieces of data that processes a model. A million tokens corresponds to about 30,000 code rules.

“Designed for integration into professional developer tools (eg intelligent codes suggestions in integrated developer environments), AI-driven coding assistants and research into code concept and generation, Mellum is also suitable for educational applications and experiments with fine tuning,” JetBrains explains in a technical report.

JetBrains says that the Mellum, which is Apache 2.0-licensed, has trained on a collection of data sets, including permissively licensed code of Github and English-speaking Wikipedia articles. Training lasted about 20 days on a cluster of 256 H200 Nvidia GPUs.

Mellum takes some work to get started. The basic model cannot be used from the box; It must first be adjusted. Although Jetbrians have provided a few Mellum models that are tailored to Python, the company warns that they are intended for “estimate about potential possibilities” not in a production environment.

AI-generated code will undoubtedly change how software is built, but it also introduces new security challenges. More than 50% of the organizations sometimes or often create safety problems with AI-produced code, according to an end of 2023 Survey by developers Security platform Synk.

JetBrains indeed notes that Mellum reflects “Prejudices present in public code bases” (for example, generating code that is comparable in style to open source repositories), and that the codes suggestions are not necessarily “secured or free of vulnerabilities”.

“This is just the beginning,” JetBrains wrote in one Blog post. “We don’t hunt for generality – we build focus. If Mellum even pops up a meaningful experiment, contribution or cooperation, we would consider it a victory.”