Leading development tool vendor JetBrains has officially released the new Mellum2.1 programming AI model, with a focus on significantly enhancing the agent programming capabilities. The model continues to use a 12B mixture of experts architecture and an open-source license, aiming to further reduce the private deployment barriers for enterprises and individual developers.

image.png

Reinforcement Learning Empowers Agent Capabilities

In terms of training mechanisms, Mellum2.1 has expanded reinforcement learning from its original short-term conclusion phase to become the core training component. It has completed specialized reinforcement in millions of sandbox environments for software engineering and algorithms. The upgraded model has strong self-retrieval and verification capabilities, accurately identifying the root cause of failed tests and drafting repair plans.

Significant Improvement in Reasoning Performance

Thanks to the introduction of multi-Token prediction technology, the model's response speed has improved by about 1.6 times in single-request scenarios. Official test data shows that when facing high-load reasoning requests, the model's Token throughput is nearly twice that of the popular open-source model Qwen3.5-9B, providing excellent cost-effectiveness for large-scale code deployment.