Mellum2: JetBrains’ 12B MoE Model for Efficient Code and Text

As AI systems grow in complexity, production pipelines increasingly rely on multiple model calls for routing, retrieval, summarization, and agent subtasks. Deploying a single frontier model for every step introduces unnecessary latency and cost. The real problem is that many of these operations are latency-sensitive and do not require the largest available model, yet most open models still overshoot on parameter count for such high-frequency tasks. JetBrains identified this tension and built Mellum2 to specifically address the need for fast, efficient inference in multi-model architectures.

Mellum2 is a 12B-parameter Mixture-of-Experts model trained from scratch on text and code, with only 2.5B active parameters per token. This design delivers more than 2x faster inference than similarly sized dense models while maintaining competitive benchmark performance on code generation, reasoning, and math. The model is released under Apache 2.0 and intentionally scoped to text and code, avoiding multimodal overhead. It targets production workflows like routing, RAG pipelines, sub-agents, and private deployments where proprietary code or data must remain on self-hosted infrastructure.

The key takeaway for builders is the strategic value of a well-scoped “focal” model. Mellum2 is not meant to replace frontier models but to make the stack faster, cheaper, and more controllable for high-frequency intermediate operations. For anyone building AI systems inside an IDE, in a RAG pipeline, or as part of an agent workflow, Mellum2 provides a pragmatic option that balances capability with deployability, demonstrating that specialization and efficiency can be more effective than brute-force scaling.