JetBrains recently announced the official open-sourcing of Mellum2, a new-generation machine learning model tailored for software engineering systems. This comes just over a year after they open-sourced their first Mellum model in 2025, which focused on code completion and had approximately 4 billion parameters. Mellum2, in contrast, expands the total parameter count to 12 billion. However, JetBrains emphasizes that by adopting a Sparse Mixture of Experts (MoE) architecture, only about 2.5 billion active parameters are actually involved in computation per token, enabling it to maintain fast inference speeds on standard hardware.
