Mellum2.1 (blog.jetbrains.com)

Mellum2.1 has been unveiled as a significant upgrade in the realm of coding agents, emphasizing its ability to work directly within repositories. This new version of the previously released 12B mixture-of-experts model boasts enhanced capabilities after extensive reinforcement learning (RL) implemented in real environments. The adoption of a new training structure has transformed RL from a mere post-training step to the core of the process, allowing Mellum2.1 to adeptly explore and manipulate codebases, edit files, and validate its own modifications. The significance of Mellum2.1 lies not only in its advanced coding capabilities but also in its efficiency, being the fastest in its class during heavy workloads, nearly doubling the token processing rate compared to similar models. It showcases improved performance across various tasks, including competitive programming and general knowledge, making it a versatile tool well-suited for private, self-hosted deployments. With its availability on platforms like Hugging Face, Mellum2.1 invites developers to leverage its capabilities in building robust AI tools and coding agents, promoting an open-source approach to enhancing model performance through community feedback.