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

JetBrains has released Mellum2, a 12B-parameter open Mixture-of-Experts (MoE) model trained on text and code. Its architecture activates only 2.5B parameters per token, enabling efficient, low-latency inference that is more than twice as fast as comparable models. Mellum2 is designed for high-throughput production workloads such as routing, RAG pipelines, and agent subtasks within larger AI systems. Released under the Apache 2.0 license, it is also suitable for private, self-hosted deployments involving proprietary data or code.