JetBrains open-sources Mellum2, a 12B coding model for software engineering
JetBrains has open-sourced Mellum2, a new machine learning model for software engineering. It upgrades from a code completion tool to a full coding assistant, expanding the context window from 8,192 to 131,072 tokens. Developers can deploy base, instruction, and reasoning versions under Apache 2.0.
Mellum2 has 12B total parameters with a sparse Mixture-of-Experts framework, activating 2.5B parameters for fast computation on standard hardware. It can generate and edit code, call external tools, run multi-step agentic workflows, and maintain long conversations.
JetBrains targets Mellum2 for AI workload routing, low-latency RAG pipelines, fast sub-agents in complex workflows, and private local AI deployment. The training used a three-phase data curriculum, moving from diverse web content to curated code and math data to better fit software engineering tasks.
Comments
Loading comments…