JetBrains open-sources Mellum2, a 12B-parameter coding model for on-prem agentic AI infrastructure.
Mellum2 fills a critical gap for enterprise engineering teams blocked by data privacy constraints from using cloud-dependent agents like Claude Code. At 12B parameters, it is small enough for cost-effective local inference while being specifically tuned for the routing and retrieval tasks that form the backbone of agentic workflows. This accelerates the shift toward fully private, self-hosted AI developer tools.
JetBrains has officially open-sourced Mellum2, a 12-billion parameter coding model designed specifically for the infrastructure layer of agentic AI systems. Unlike consumer-facing coding assistants or cloud-dependent agents like Anthropic's Claude Code, Mellum2 is built to run entirely on private infrastructure.
Technical Details At 12B parameters, Mellum2 hits a sweet spot for local deployment, requiring significantly less VRAM than massive frontier models while maintaining high capability in specialized domains. The model is specifically optimized for infrastructure-level tasks within agentic systems. Rather than just generating boilerplate code, it is tuned for complex routing, managing retrieval-augmented generation (RAG) pipelines, and executing sub-agent tasks. This makes it a foundational component for developers building multi-agent architectures using modern frameworks like Mastra.
Why It Matters From an engineering perspective, the reliance on third-party APIs for AI coding tools has been a massive bottleneck for enterprises with strict data privacy and compliance requirements. Tools that send proprietary codebase context to external servers are often non-starters for regulated industries. Mellum2 provides a highly capable, open-source alternative that runs entirely on-premises. By targeting the "glue" layer of AI agents—routing and retrieval—JetBrains is commoditizing the underlying orchestration logic needed to build autonomous coding systems. It allows engineering teams to build custom, self-hosted AI agents that have full access to internal codebases, databases, and documentation without leaking IP to third parties.
What to Watch Next Keep an eye on how the open-source community integrates Mellum2 into existing agent frameworks. We should expect to see a surge in privacy-first, locally hosted coding assistants that rival cloud-based offerings for enterprise use cases. Additionally, monitor JetBrains' own ecosystem; their native IDE integrations will likely leverage this model to offer hybrid local/cloud AI capabilities, setting a new standard for secure enterprise developer environments.