JetBrains open-sources Mellum 2.1 coding agent model on October 8 under Apache 2.0 on Hugging Face

JetBrains released Mellum2.1 on October 8, the company's open-source large language model for coding agents. The model has 12B total parameters and roughly 2.5B active parameters, and the weights are published on Hugging Face under Apache 2.0. JetBrains says Mellum2.1 was trained with reinforcement learning on real code repositories, which the company credits for improvements in code exploration, file editing, and change verification, with special attention to code completion, long-file editing, and function-level refactoring scenarios. Mellum2.1 is the latest in the Mellum series and reaches roughly twice the throughput of Qwen3.5-9B on high-load inference benchmarks in published tests, serving as the foundation for JetBrains' gradual introduction of local LLM support in its IDE products and in JetBrains AI Assistant, with on-device agent workflows coming next. JetBrains and Hugging Face co-published the release so enterprise teams can deploy it on infrastructure they control, meeting requirements that code snippets must not leave the internal environment.