JetBrains has open-sourced Mellum, a specialized 4-billion-parameter language model tailored for software development tasks.
Mellum is optimized for programming-related tasks like autocompletion and structural understanding of source code in various languages.
The model was trained using over 4.2 trillion tokens and achieves strong performance in benchmarks, reflecting its focus on structured code understanding.
JetBrains released Mellum under the Apache 2.0 license to promote transparency, reusability, community collaboration, and pedagogical value, indicating a shift toward specialized, efficient language models for developer tooling.