JetBrains releases open code model Mellum2.1 with 12 billion parameters

JetBrains has released Mellum2.1, an open AI model aimed at coding agents. That's according to a headline and a short description from MarkTechPost, which is the only source this time.

The model has 12 billion parameters and is built on a Mixture of Experts architecture, abbreviated MoE. In this type of model, the network is split into several specialized parts, so-called experts, and typically only some of them are active at a time. It is a common way to keep computing power down relative to the model's total size.

The description calls the model open, but the sources say nothing about exactly which license applies or how open the weights and training data are.

What it means in everyday life

The target group is coding agents, that is, AI tools that do not just suggest code but can solve tasks in multiple steps on their own. For developers and companies in Denmark, an open model can be interesting because it in principle allows you to choose your own setup rather than only using closed services.

However, we do not know details about performance, benchmarks, supported programming languages, hardware requirements or price. It is therefore too early to assess how Mellum2.1 compares with other models. If you are considering using it, you should wait for the full coverage and JetBrains' own documentation.

What it means for you

If you or your company have developers working with code, an open model like Mellum2.1 can give more freedom to choose how the AI help is set up, rather than only using closed services. It is most relevant for those who want to build or try out AI tools that solve tasks in multiple steps on their own. If you are not a developer yourself, there is nothing you need to do now, but it is a good sign that more coding tools will be built on open models. If you want to try it, go to JetBrains' own documentation before you decide.