JetBrains releases Mellum2, an open 12B MoE model for fast tasks in AI pipelines

JetBrains, the company behind the IntelliJ IDEA and PyCharm IDEs used by tens of millions of developers, has released Mellum2, an updated version of its specialized language model. Mellum2 is available under the Apache 2.0 license and is built for fast handling of specialized tasks within multi-model AI pipelines. Mellum is JetBrains' own family of language models for programming tasks. Mellum2 is a significant update: a new MoE architecture, a larger training dataset and clear positioning as a specialized component for multi-model systems.

MoE architecture and training scale

Mellum2 is built on a Mixture of Experts (MoE) architecture with 12 billion parameters. Unlike classic dense models, where every weight in the network is activated for each request, MoE splits the model into specialized "expert" blocks and activates only some of them depending on the token type. This lets the model run faster and more cheaply at the same capacity.

For pretraining, JetBrains used a corpus of 10.6 trillion tokens, a scale comparable to flagship open models such as LLaMA 3 from Meta*. For a coding model, data volume matters a great deal: the wider the coverage of programming languages, frameworks and technical documentation, the more accurately the model handles tasks in real codebases.

Role in multi-model pipelines

Mellum2 does not aim to compete with general-purpose assistants such as Claude or GPT. It was designed as a fast, specialized component for specific steps in a developer's workflow:

• real-time code completion directly in the IDE
• fast refactoring of small functions and blocks of code
• request classification and routing inside AI agents
• generation of comments and inline documentation
• "cheap" intermediate steps before a task is handed to a larger model

The last scenario is the key one. Production AI systems for development are increasingly built as "orchestras of models": a light, fast model handles routine requests with minimal latency, while a heavy flagship model is called in only for complex tasks that require deep reasoning. Mellum2 is built for exactly the role of fast specialist in such pipelines.

Open Apache 2.0 license

Apache 2.0 is one of the most permissive open licenses. Mellum2 can be embedded in closed corporate products, fine-tuned on proprietary data and used in commercial services without paying royalties to JetBrains. The company is publishing the model weights openly.

Open-weight coding models already face notable competition: Code Llama from Meta*, DeepSeek-Coder, StarCoder, Qwen2.5-Coder. Mellum2 enters this field with a specific positioning: deep integration into the JetBrains IDE ecosystem and a role as a specialized component in multi-stage pipelines, rather than an attempt to beat the flagships on general code-generation benchmarks.

What this means

Mellum2 points to an important trend: companies that build developer tools are ceasing to buy AI from outside providers and are starting to train their own specialized models for their users' specific scenarios. JetBrains is betting that deep IDE integration and domain specialization matter more than abstract superiority on general-purpose benchmarks.

For developers, this means faster AI suggestions right in their familiar editor, without cloud API latency and without having to send their own code to third-party servers.

*Meta has been designated an extremist organization and banned in Russia.