Introducing Mellum2: A 12B Mixture-of-Experts Model by JetBrains

JetBrains has introduced Mellum2, an open Mixture-of-Experts model featuring 12 billion parameters and optimized specifically for low-latency text-and-code workloads. Building on the foundation of the original Mellum code completion model, Mellum2 expands its capabilities to encompass a broader set of software engineering and natural language tasks while prioritizing efficient inference and deployability. According to the release details, modern AI systems frequently rely on multiple latency-sensitive model calls such as routing, retrieval, summarization, planning, validation, and tool use, which do not necessitate the largest available models. Mellum2 is engineered to target these specific workloads. In technical evaluations covering code generation, reasoning, science, and math benchmarks, Mellum2 is competitive with similarly sized open models while offering more than twice the inference speed. The Mixture-of-Experts architecture maintains high total model capacity while activating only a subset of parameters per token, lowering serving costs for real-time operations. By focusing exclusively on text and code rather than multimodal tasks, the model remains compact. JetBrains envisions Mellum2 acting as a focal model within larger AI architectures, functioning effectively as a lightweight router, an orchestration tool for multi-model systems, a component in latency-sensitive retrieval pipelines, and a utility for agent subtasks. Because it is open and efficient, developers can deploy Mellum2 in self-hosted environments to handle internal data and proprietary code securely, helping make production AI stacks faster, cheaper, and easier to control.