Software development tool maker JetBrains has released Mellum2.1, a new open-source model built specifically for coding agents. According to a report by MarkTechPost, the model uses a mixture-of-experts architecture with a total of 12 billion parameters, of which 2.5 billion are active during inference. It is distributed under the Apache 2.0 license. The developers applied reinforcement learning directly within real software repositories, a process that reportedly increased its SWE-bench Verified benchmark score from 2.0 to 47.0.
The Architecture of Coding Agents
Mixture-of-experts models divide tasks among different sub-networks, or experts, activating only a subset of parameters for any given token or prompt. This design aims to balance high capacity with lower computational costs during deployment. For coding assistants and autonomous agents, fine-tuning on actual codebases via reinforcement learning is a common method to improve syntax correctness, repository navigation, and bug resolution capabilities. The open licensing of Mellum2.1 allows developers and researchers to inspect, modify, and deploy the weights locally or within custom pipelines.
Unanswered Questions and Technical Details
Despite the release, several operational specifics remain unknown. The announcement did not specify the exact hardware requirements needed to run or fine-tune Mellum2.1 efficiently, nor did it detail the composition and size of the training dataset used alongside the reinforcement learning phase. Additionally, performance benchmarks beyond SWE-bench Verified were not provided in the initial disclosure, leaving open questions about its latency, multi-language proficiency, and behavior in enterprise-scale environments.
Developers interested in testing Mellum2.1 can access the model files under its Apache 2.0 terms, though further real-world evaluation will be necessary to gauge its practical utility across diverse development workflows.
Lead image: Pixabay / Pexels
