Mellum2 is a software engineering workflow-specific language model released by JetBrains on June 1, 2026. It utilizes a Mixture-of-Experts (MoE) architecture with a total of 12B parameters, capable of processing both natural language and source code. During inference, only approximately 2.5B parameters are activated per token. This design is similar to deploying only a small number of specialized teams that match the nature of the request, rather than invoking the entire large development organization each time. JetBrains describes Mellum2 as a fast base model that can be used in various stages of AI development workflows, such as request routing, question answering, summarization, and sub-agent tasks, rather than simply a code generation model. Typical large models can handle complex problems broadly, but using large computational resources for every request can quickly increase latency and inference costs in coding agents with frequent invocations. The key differentiator of Mellum2 is that it does not always compute all 12B parameters; instead, it activates only a portion of the routed experts, aiming to balance the throughput and response speed required for software engineering tasks. In particular, in an agent environment where a single user request cascades into intent classification, repository context summarization, question answering, and sub-task assignment, both the peak performance of the model and the cost and latency per invocation are important. However, the official evaluation benchmarks, supported context length, memory usage per precision, and performance metrics compared to other coding models cannot be confirmed based solely on the provided Discovery information, requiring further validation. Another key aspect is its local execution and self-hosting capabilities. Teams that handle private repositories that are difficult to send to external APIs can deploy Mellum2 in their internal execution environment and consider a configuration where requests containing code and natural language are processed within the organization's control. For example, a top-level agent can classify incoming development requests, then use Mellum2 to summarize related modules and issue context, and route them to sub-tasks such as test writing or document updates. The results generated by the model are best used by passing them through static analysis, unit tests, and code reviews, rather than merging them immediately. The supported inference frameworks, quantization formats, and actual deployment commands require checking the official model card. For life science software researchers, it is more meaningful as a component that assists research code workflows rather than a model that replaces the knowledge of published papers. In a Python-based analysis repository, it can be used to summarize change requests or distribute sub-tasks by providing data preprocessing functions, experiment configuration files, test code, and README files together as context. It can also be evaluated as a pipeline documentation and question-answering assistant in a self-hosted environment that does not send sensitive research code to external services. However, bioinformatics accuracy, suitability for medical purposes, or compatibility with specific frameworks have not been confirmed, so it should not be used to replace research result validation and domain expert review.

Mellum2 is a software engineering workflow-specific language model released by JetBrains on June 1, 2026. It uses a Mixture-of-Experts (MoE) architecture with a total of 12B parameters, capable of processing both natural language and source code. During inference, only approximately 2.5B parameters are activated per token. This design is similar to deploying only a small number of specialized teams that match the nature of the request, rather than calling the entire large development organization each time. JetBrains positions Mellum2 as being more than just a simple code generation model; it is designed for request routing, question answering, summarization, and sub-agent tasks.