JetBrains open sources Mellum, its specialized AI model for code completion

JetBrains has announced that it is open sourcing its Mellum language model, making a base version of the code-focused AI available on Hugging Face.

If you don't know what Mellum is, it is JetBrains' own specialized large language model built from the ground up specifically for helping developers. It is designed to be good at code-related tasks, primarily focused on making code completion much faster, more accurate, and smarter within JetBrains IDEs like IntelliJ IDEA, PyCharm, and others.

It is different from bigger, general-purpose AI models because it is built for a specific job, not trying to do everything at once. The company says this approach helps it provide suggestions much quicker and with better relevance to the code you are actually writing. It reported significant improvements in their AI Assistant's code completion speed and accuracy when using Mellum compared to previous implementations.

So why would JetBrains open source a model that seems so central to its own paid AI features? Well, according to JetBrains, it comes down to a belief in transparency and collaboration. JetBrains points out that open source projects like Linux and Git have driven huge progress in tech, and believes open source AI could do the same. By putting the base Mellum model out there, JetBrains hopes to give researchers, educators, and advanced teams a look under the hood of a model built specifically for code tasks.

JetBrains refers to Mellum as a "focal model". This concept involves building an AI that excels at a single, specific task instead of trying to handle a wide range of functions like a generalist model. Supporters of this approach point to advantages such as greater precision for targeted tasks, lower operational costs, a reduced environmental footprint, and improved accessibility for researchers and smaller teams without the resources required for large-scale general models.

Now, the version of Mellum available on Hugging Face is the 4-billion parameter base model. It is multilingual and optimized for code completion. JetBrains shared some benchmark data comparing Mellum to some other code models:

The table shows how Mellum stacks up on different coding benchmarks. It is not necessarily outperforming the larger CodeLlama models in every test, but it shows solid performance for its size and specialized nature, especially considering its focus on speed and efficiency for integrated code completion.

JetBrains is clear that this open-sourced base model is likely not something the average developer will just download and run for their daily coding. Instead, it is aimed more at AI and machine learning researchers looking into code AI, engineers and educators who want to learn about or teach building domain-specific language models, or advanced teams interested in adapting such models.

This is just the beginning for Mellum, according to JetBrains. The company plans for it to grow into a family of focal models, each built for different specific coding tasks beyond just completion, perhaps even predicting code changes or diffs.

In other JetBrains AI news, the company recently updated its AI Assistant offering. They introduced a new free tier, allowing more developers to try out some of the AI features without a subscription, albeit with limits on usage. Crucially, the AI Assistant also added support for more AI models beyond just OpenAI, including Google Gemini. This gives users more options for which AI model powers their coding assistance within the IDE.