Mellum2発表：JetBrainsによる120億パラメータのMixture-of-ExpertsモデルIntroducing Mellum2: A 12B Mixture-of-Experts Model by JetBrains

JetBrains has released Mellum2, a 12-billion-parameter language model built on a Mixture-of-Experts (MoE) architecture and tuned specifically for code completion. The model is notable because it comes from a company best known for developer tooling rather than foundation-model research, and because it targets a practical niche—assisting programmers inside an integrated development environment—rather than competing as a general-purpose chatbot. For teams interested in running code assistance locally or on their own infrastructure, an openly available model of this kind is worth attention.

Mellum2 follows the original Mellum, which JetBrains positioned as a "focal" model: a comparatively small system trained to do one job, code completion, well, rather than a large model expected to handle every task. The second generation appears to extend that philosophy while significantly increasing capability. Moving to a Mixture-of-Experts design is the headline architectural change. In an MoE model, the network is divided into multiple expert subnetworks, and a routing mechanism activates only a subset of them for each token processed. The practical benefit is that the model can hold a large total parameter count—useful for capacity and knowledge—while keeping the number of parameters actually used per inference step relatively low. That combination is attractive for code completion, where latency matters because suggestions must appear as the developer types.

The 12B figure refers to the total parameter budget; MoE models typically activate only a fraction of their weights at once, so the effective compute per token is generally lower than the headline number suggests. JetBrains has not framed Mellum2 as a replacement for large frontier models, and the design choices are consistent with an emphasis on responsiveness and deployment efficiency rather than maximal reasoning breadth. This is a familiar trade-off in the code-model space, where the goal is fast, contextually relevant completions across many programming languages and file types.

Distribution through Hugging Face places Mellum2 alongside a growing catalog of open or openly licensed code models. Comparable efforts include Meta's Code Llama, the StarCoder models from the BigCode project, DeepSeek's Coder series, and Alibaba's Qwen Coder line. Each targets similar workflows—completion, infilling, and sometimes instruction-following for code tasks—and the field has moved steadily toward smaller, specialized models that can run on modest hardware. Mellum2's MoE approach fits this trend of squeezing more capability into deployable footprints, and its release adds another option for developers who prefer not to depend on hosted, proprietary services such as GitHub Copilot.

The strategic context is also relevant. JetBrains ships widely used IDEs, including IntelliJ IDEA, PyCharm, WebStorm, and others, and has an existing AI assistant offering. A model optimized for code completion aligns naturally with that product portfolio, and it is reasonable to expect Mellum2 or its successors to inform in-editor features. Owning the underlying model gives a tooling vendor more control over cost, latency, privacy, and the ability to fine-tune behavior for specific languages or codebases—factors that matter to enterprise customers who are cautious about sending source code to external APIs. Whether the openly released weights are intended primarily as a research artifact, a community resource, or a foundation for commercial features is not fully clear from the announcement alone, and prospective users should check the accompanying license terms before integrating it into products.

For practitioners, a few points are worth verifying directly from the model card. These include the exact number of active parameters per token, the context window length, the set of supported programming languages, the training data and any stated benchmarks, and the license governing commercial use. Code completion quality is difficult to judge from headline specifications, and real-world usefulness depends heavily on how well the model handles a given team's languages and coding conventions.

Overall, Mellum2 represents an incremental but meaningful step in the specialization of code models: a purpose-built assistant, an efficiency-oriented MoE architecture, and open availability aimed at IDE integration. Its ultimate impact will likely depend on measured performance against established code models and on how smoothly it can be embedded into everyday development workflows.