JetBrains, a well-known provider of development tools, has introduced its latest innovation—the Mellum2.1 programming AI model. This model is specifically designed to enhance agent programming capabilities. Building on the successful 12B Mixture of Experts (MoE) architecture, Mellum2.1 is now available under an open-source license, making it easier for enterprises and individual developers to deploy privately. One of the key advancements in Mellum2.1 is its elevation of reinforcement learning from a brief finishing touch to a fundamental aspect of the training process. After undergoing specialized reinforcement training in software engineering and algorithms across millions of sandbox environments, Mellum2.1 has developed robust autonomous retrieval and verification capabilities. This enables it to precisely identify the root causes of failed tests and propose effective repair solutions. Additionally, with the integration of multi-token prediction technology, the model's response speed per request has increased by approximately 1.6 times. Under high loads, its token throughput is nearly double that of the similar and widely-used open-source model, Qwen3.5-9B. This significant improvement in inference performance offers a cost-effective solution for large-scale code deployment.