JetBrains has released Mellum2, a new 12-billion parameter Mixture-of-Experts model designed specifically for low-latency text and code generation. Unlike traditional dense models, Mellum2 activates only 2.5 billion parameters per token during inference, offering a significant improvement in speed and efficiency for high-throughput applications. Originally developed as a code completion tool, the model has evolved to handle a broader spectrum of natural language and software engineering tasks. This expansion maintains the project's core focus on efficiency and deployability. By specializing in text and code rather than attempting multimodal capabilities, Mellum2 remains compact and optimized for the specific demands of software engineering workflows. Performance benchmarks indicate that Mellum2 delivers competitive results against other models of similar size while achieving more than twice the inference speed. This performance makes it particularly suitable for production environments where response time is critical. The model is licensed under the Apache 2.0 agreement and is available for download via Hugging Face. The architecture leverages a Mixture-of-Experts design, allowing the system to maintain high total capacity while routing only a small subset of experts to process each input token. This approach reduces serving costs and latency, addressing the needs of modern AI systems that rely on multiple sequential model calls rather than a single monolithic model. Key use cases for Mellum2 include routing and orchestration within complex systems. It can efficiently handle prompt classification, tool selection, and control flow decisions. In Retrieval-Augmented Generation pipelines, the model is well-suited for latency-sensitive tasks such as context compression, summarization, and post-processing retrieved information. Additionally, it functions effectively as a sub-agent for planning, validation, and context preparation, reducing the reliance on larger, slower models for intermediate operations. The release also supports private deployments. Because the model is open-source and efficient to serve, organizations can host it on self-managed infrastructure to ensure the security of proprietary code and internal data. JetBrains envisions Mellum2 as a focal component within a larger, specialized AI stack. Rather than replacing large reasoning models, it aims to accelerate the entire system by handling high-frequency, specific tasks. This modular approach seeks to make AI pipelines faster, more cost-effective, and easier to manage. For developers building software engineering tools, IDE integrations, RAG pipelines, or agent workflows, Mellum2 offers a ready-to-use solution. Technical details regarding the architecture, training setup, and evaluation methodology are documented in a full report available on arXiv. As AI systems mature, the industry is shifting toward specialized architectures where well-scoped models like Mellum2 play a vital role in optimizing performance across diverse operational requirements.