JetBrains releases Mellum, an ‘open’ AI coding model

Jetbrains, a company behind a series of popular applications development tools, has released its first “Open” AI model for coding.

They made Jetbrains on Wednesday MeasureA model that generates the code published by the company for various software apartments last year, openly available on AI DeV platform hug. Mellum, dressed in more than 4 trillion of token, weighs 4 billion parameters, and is designed especially for the end of the code (ie, the end of the code clips based on the surrounding context).

The parameters are approximately corresponding to the skills of solving the model problems, while the tokens are raw bits of the data that the model processes. One million tokens are equivalent to ~ 30,000 lines of code.

“Designed for integration into professional development tools (eg intelligent code in integrated environmental environments for development developers), assistants of coding on AI facility and exploration of the understanding and generation of code, Mellum is also suitable for educational applications and experiments in precise settings,” ” explains Jetbrains in a technical report.

Jetbrains says he trained Mellum, who was Apache 2.0-boiled, collecting data sets, including a permitted licensed code from Github and Wikipedia articles in English. The training lasted about 20 days on the 256 H200 cluster NVIDIA GPU.

Mellum takes a job to get up and run. The basic model cannot be used outside the frame; First must be finely adjusted. Although Jetbrians provided several mellum models that are precise for Python, the company warns that they are intended for “assessment of potential abilities” without being deployed to the production environment.

Ai-gearred code no doubt changes the way the software is built, but also represents new security challenges. More than 50% of organizations are encountered with safety problems with a-based code sometimes or often, towards the end of 2023 Synk Programming Safety Player.

Indeed, Jetbrains notes that Mellum can “reflect the bias present in public codes” (eg, a code that generates in a style of open code), and that its code suggestions will not necessarily be “safe or without vulnerability.”

“This is just the beginning,” Jetbrains ua wrote blog blog. “We do not persecute generality – we are building a focus. If Mellum shows even one meaningful experiment, contribution or cooperation, we would consider it a victory.”