Jetbrains, the company behind a range of popular application development tools, has launched its first “open” model for coding.
On Wednesday, Jetbrains made Mellum, a code generating model that the company launched for its various software development suites last year, openly notable on the AI dev Hugging Face platform. Mellum, trained in more than 4 billion tokens, weighs 4 billion parameters and is specifically designed for the complete code (the ending code fragments based on the surrounding context).
The parameters correspond approximately to the problem solving skills of a model, while the tokens are unprocessed data bits that process a model. One million tokens corresponds to approximately 30,000 lines of code.
“Designed for integration into professional developer tools (Smart Code Suggestions in Integrated Developer Envidents), Coding assistants with AI food and research on understanding and code generation, Mellum is also a good adjustment in adjustment A,” explanations “explanations”, explanations “, explanations”, explanations. ” Explanations, Report.
Jetbrains says he trained Mellum, who has an Apache 2.0 license, in a collection of data sets that include the code with permanent license of Wikipedia articles in the English language. The training took about 20 days in a group or 256 GPU H200 NVIDIA.
Mellum requires some work to put into operation. The base model cannot be used outside the box; First it has to be adjusted. While Jetbrians has tried some tight Mellum models for Python, the company warns that they are intended for “estimation on potential capacities”, not being implemented in a production environment.
The code generated by AI is undoubtedly changing how the software is built, but is also introducing new security challenges. More than 50% of organizations encounter security problems with the code produced by AI-AI to some or often, according to a survey at the end of 2023 by the Sink of the developer’s security platform.
Techcrunch event
Berkeley, ca.
|
June 5
Book now
In fact, Jetbrains points out that Mellum can “reflect the biases present in the public code bases” (for example, that generates similar code in style to open source repositories), and that their code suggestions necessarily won “safe or free of vulnerabilities.”
“This is just the beginning,” Jetbrains wrote in a blog post. “We are not chasing the generality, we are building approach. If Mellum causes an equally significant experiment, contribution or collaboration, we would like me to come.”
