JetBrains released Mellum2 — an open 12B MoE model for fast tasks in AI pipelines

JetBrains — the company that created IDE tools IntelliJ IDEA and PyCharm used by tens of millions of developers — has released Mellum2, an updated version of its specialized language model. Mellum2 is released under the Apache 2.0 license and is designed for fast processing of specialized tasks as part of multi-model AI pipelines.

Mellum is JetBrains' proprietary line of language models for programming tasks. Mellum2 is a significant update: new MoE architecture, scaled training data volume, and clear positioning as a specialized component for multi-model systems.

MoE Architecture and Training Scale

Mellum2 is built on Mixture of Experts (MoE) architecture with 12 billion parameters. Unlike classical dense models where all network weights are activated with each request, MoE divides the model into specialized "expert" blocks and activates only a portion of them depending on token type. This allows the model to operate faster and more economically with the same capacity.

For pretraining, JetBrains used a corpus of 10.6 trillion tokens — a scale comparable to flagship open models like LLaMA 3 from Meta. For a coding model, data volume is critically important: the wider the coverage of programming languages, frameworks, and technical documentation, the more accurately the model handles tasks in real code bases.

Role in Multi-Model Pipelines

Mellum2 does not aspire to compete with universal assistants like Claude or GPT. It was designed as a fast specialized component for specific steps in developer workflows:

• real-time code autocomplete directly in the IDE
• fast refactoring of small functions and code blocks
• query classification and routing within AI agents
• generation of comments and inline documentation
• execution of "cheap" intermediate steps before passing a task to a larger model

The last scenario is key. Production AI systems for development are increasingly built as "model orchestras": a light fast model processes routine requests with minimal latency, a heavy flagship model is only engaged for complex tasks requiring deep reasoning. Mellum2 was created precisely for the role of fast specialist in such pipelines.

Open Source Apache 2.0 License

Apache 2.0 is one of the most liberal open source licenses. Mellum2 can be embedded in closed corporate products, fine-tuned on proprietary data, and used in commercial services without royalty payments to JetBrains. The company publishes model weights in open access.

Among coding models with open weights there is significant competition: Code Llama from Meta, DeepSeek-Coder, StarCoder, Qwen2.5-Coder. Mellum2 enters this space with specific positioning — deep integration into the JetBrains IDE ecosystem and the role of a specialized component in multi-stage pipelines, not an attempt to outpace flagships in general code generation benchmarks.

What It Means

Mellum2 is evidence of an important trend: companies creating tools for developers are moving away from buying AI from external providers and beginning to train their own specialized models for specific user scenarios. JetBrains is betting that deep IDE integration and domain specialization matter more than abstract superiority in universal benchmarks. For the developer this means faster AI suggestions directly in the familiar editor — without cloud API delays and without the need to send own code to third-party servers.