Mellum2 gives workplace tools a compact model component

JetBrains’ Nikita Pavlichenko introduced Mellum2, a mixture-of-experts model with 12 billion total parameters and a smaller active portion per token. The publication identifies routing, retrieval, summarization, and supporting-agent uses. Weights are available under Apache 2.0.

Context

This kind of model is interesting as a component rather than an all-purpose conversational destination. A fast, bounded operation can improve the larger product. Savings still depend on how difficult cases reach a stronger model and how failures are detected at the boundary between components.