JetBrains has released Mellum2, a 12B-parameter mixture-of-experts (MoE) model, as detailed in a Hugging Face blog post. The model is designed for code completion, a core feature in JetBrains' family of IDEs. By using an MoE architecture, Mellum2 aims to balance model capacity with computational efficiency, activating only a subset of parameters per inference step.
The post positions Mellum2 as a step forward in code-specific language modeling, though it does not provide benchmark comparisons or details on training data. The focus is on the architectural choice and the model's intended use case rather than on broader coding-agent capabilities.
Since this is the only source, there is no independent verification of the model's performance claims. The announcement is primarily a technical introduction, leaving questions about real-world effectiveness open for future evaluation.