JetBrains Releases Mellum2: 12B Parameter Mixture-of-Experts Architecture Developer-Focused Model

Short summary

JetBrains released Mellum2, a 12B parameter open-source MoE model that activates only 2.5B parameters per inference, doubling speed versus equivalent-scale models while cutting deployment costs. It targets high-frequency lightweight tasks—prompt classification, tool selection, RAG context compression, and code completion—within multi-model collaboration architectures rather than competing with frontier models. Its Apache 2.0 license and lean text/code-only design make it a pragmatic choice for enterprises running AI in private environments.

• •JetBrains open-sourced Mellum2: 12B total params, 2.5B active via MoE, Apache 2.0 license
• •Designed for lightweight tasks in multi-model systems: prompt classification, tool selection, RAG compression, code completion
• •Practical guidance: audit your multi-model pipeline nodes and replace non-critical positions with focused models to cut inference costs

Generated with AI, which can make mistakes.