JetBrains open-sourced Mellum 2 on June 5 — a 12B sparse MoE coding model. The "focal model" approach is to have a smaller, focused model do the bulk of the coding work, with a larger, more general model invoked only for complex cases, splitting the cost-quality trade-off. On JetBrains' internal production chain, Mellum 2 covers code completion, code review, and test generation, with significant quality improvements over Mellum 1.