Details
- JetBrains released Mellum2, a 12B-parameter Mixture-of-Experts model trained from scratch on natural language and code, with only 2.5B parameters active per token for efficient inference.
- The model is developed by JetBrains as a successor to the original Mellum code completion model, and is available on Hugging Face under the Apache 2.0 open-source license.
- Mellum2 targets latency-sensitive tasks such as routing, RAG, summarization, sub-agents, high-throughput coding features, and private/self-hosted deployments, emphasizing low-latency, high-throughput text-and-code workloads.
- Compared with similarly sized open models, Mellum2 is reported to deliver competitive benchmark performance while achieving more than 2x faster inference, helped by its MoE architecture that activates only a subset of experts per token.
- The technical report describes base, instruct, and thinking variants specialized for software engineering tasks like code generation, editing, debugging, tool use, and agentic workflows, positioning Mellum2 as a compact focal model within multi-model AI systems.
Impact
Mellum2 strengthens the emerging pattern of multi-model AI stacks where smaller, specialized models handle high-frequency orchestration, routing, and code-centric tasks while larger frontier models focus on heavy reasoning. Its open Apache 2.0 license and MoE efficiency are likely to accelerate self-hosted, IDE-integrated, and agentic coding deployments over the next 12–24 months, and may push competitors to release similarly efficient, code-focused MoE models for production pipelines.
