JetBrains introduces Mellum2, a fast 12B AI model that excels in software engineering

A year after introducing its first open AI code generation model, JetBrains has released Mellum2, a 12-billion parameter AI model focused on code generation and efficiency in production environments. Developed from scratch and distributed under the Apache 2.0 open-source license, Mellum2 offers organizations a cost-effective alternative to proprietary large language models.

While building on the original Mellum model's code completion foundation, Mellum2 extends its capabilities to support both natural language and code. This allows the model to assist with routing, summarization, orchestration, and intermediate reasoning tasks within modern artificial intelligence workflows.

The model's architecture uses a Mixture-of-Experts (MoE) design, activating only 2.5 billion parameters per token. This approach reduces computing requirements and supports high-throughput, low-latency inference, making Mellum2 suitable for real-time applications at scale. Mellum2 is not multimodal and has been trained solely on natural language and code datasets. As a result, it is optimized for software engineering environments, excelling where performance and specialization are needed.

Among Mellum2's key use cases are orchestrating artificial intelligence workloads, powering fast retrieval-augmented generation (RAG) pipelines, managing sub-agents in complex workflows, and supporting private, on-premises deployments.