JetBrains has introduced Mellum, its proprietary AI model tailored for software development tasks.
Mellum is exclusively integrated into JetBrains’ AI Assistant, delivering significant enhancements in both the speed and accuracy of code completions compared to earlier versions.
In contrast to larger language models, Mellum is designed with a smaller footprint to provide near-instant coding suggestions. Currently, it supports five major programming languages—Java, Kotlin, Python, Go, and PHP—with plans for additional languages through early access.
Since Mellum’s integration, performance metrics have shown impressive results. JetBrains reports that completion latency has decreased to one-third of its previous duration, and the acceptance rate for suggestions has reached around 40%, setting a strong industry benchmark.
The company has also noted a three to four-fold drop in suggestion cancellations, along with more than doubling the total number of completions offered to developers. JetBrains emphasizes its commitment to privacy, stating that Mellum has been trained solely on publicly available code with permissive licensing, and it neither stores nor uses customer inputs during operation.
Vladislav Tankov, Director of AI at JetBrains, commented:
“The JetBrains AI Assistant intelligently selects the best model from OpenAI and Google for each use case. However, we realized that to achieve truly powerful code completion, we needed to enhance the Assistant with our own model. The launch of the Mellum model represents a significant advancement for us.
Moreover, the true strength of our code completion system lies in Mellum’s deep integration with JetBrains IDEs. The collaboration between client-side integration and server-side logic is essential for providing fast, accurate, and context-aware suggestions, ensuring an improved coding experience with remarkable results.”
By Shadjavist.
Author: shadjavist
Humble gentleman who never believes in impossibility
