Mellum Goes Open Source: A Purpose-Built LLM for Developers, Now on Hugging Face | The JetBrains Blog (2025)

Supercharge your tools with AI-powered features inside many JetBrains products

Follow

  • Follow:
  • RSS RSS

Explore More

JetBrains AI

Read this post in other languages:

  • Français

Mellum doesn’t try to know everything. It’s designed to do one thing really well: code completion. We call it a focal model – built with purposeful depth and not concerned with chasing breadth.

But code completion is just the start.

Mellum will grow into a family of focal models, each specialized for different coding tasks – from code completion to diff prediction and beyond.

Now, the base model is open-sourced and available on Hugging Face. Whether you’re building tools, running research experiments, or just curious, you’ll have full access to a fast, multilingual model*.

*Mellum supports code completion for Java, Kotlin, Python, Go, PHP, C, C++, C#, JavaScript, TypeScript, CSS, HTML, Rust, Ruby.

🤔 Why open-source Mellum?

This question was the subject of a big internal discussion.

Mellum is not just a fine-tuned version of an open-source model. It’s a model we trained from scratch to power cloud-based code completion in JetBrains IDEs, and it was released to the public last year. It’s also the first in a planned family of code-specialized models.

So why open-source it?

Because we believe in transparency, collaboration, and the power of shared progress. From Linux and Git to Node.js and Docker, open source has driven some of the biggest leaps in technology. With open-source LLMs now outperforming some industry leaders, it’s reasonable to assume that AI’s general evolution might follow a similar trajectory.

Mellum isn’t a plug-and-play solution. By releasing it on Hugging Face, we are offering researchers, educators, and advanced teams the opportunity to explore how a purpose-built model works under the hood.

What is a focal model?

In machine learning, specialization isn’t new – it’s a core approach that has guided model design for decades, with models built to solve specific tasks efficiently and effectively. Somewhere along the way, the AI conversation shifted towards general-purpose models that aim to do everything, often at a massive computational and environmental cost.

Focal models return to that original principle: build models to excel in one area.

Think of it like T-shaped skills – a concept where a person has a broad understanding across many topics (the horizontal top bar or their breadth of knowledge), but deep expertise in one specific area (the vertical stem or depth). Focal models follow this same idea: they aren’t built to handle everything. Instead, they specialize and excel at a single task where depth truly delivers value.

Mellum is a clear example. It’s a small, efficient model designed specifically for code-related tasks, starting with code completion.

Why did we take this approach? Because not every problem demands a generalist solution, and not every team has the resources or need to run large, catch-all models.

Focal models like Mellum offer clear advantages:

  • Purpose-built precision for domain-specific tasks
  • Cost efficiency when it comes to running and deploying them
  • Lower computation and carbon footprints
  • Greater accessibility for researchers, educators, and smaller teams

This isn’t a step backward – it’s applying proven principles of specialization to modern AI problems. We see that as a smarter way forward.

How does Mellum perform?

Mellum is a multilingual, 4B parameter model optimized specifically for code completion. We benchmarked it on several datasets across multiple languages, and also ran extensive human evaluations in JetBrains IDEs. In this post, we’ll provide some evaluation data for Mellum compared to some bigger models. Full details, results, and comparisons are available on the model card.

HumanEval InfillingRepoBench 1.1 (2K context, py)SAFIM (avg)
single-linemulti-line
Mellum-4B-base66.238.528.238.1
InCoder-6B69.038.633.8
CodeLlama-7B-base83.050.834.145.0
CodeLlama-13B-base85.656.136.252.8
DeepSeek-Coder-6.7B80.763.4

Who Mellum is (and isn’t) for

Let’s be real – the average developer probably won’t fine-tune or deploy Mellum. That’s okay.

Instead, the current version of Mellum on Hugging Face is meant for:

  1. AI/ML researchers: Especially those exploring AI’s role in software development, benchmarking, or model interpretability.
  2. AI/ML engineers and educators: As a foundation for learning how to build, fine-tune, and adapt domain-specific language models, or to support educational programs focused on LLM architecture and specialization.

Try Mellum today

Mellum is now live on Hugging Face. This is just the beginning. We’re not chasing generality – we’re building focus. If Mellum sparks even one meaningful experiment, contribution, or collaboration, we would consider it a win.

We’d love for you to join us by trying Mellum for yourself.

Explore Mellum on Hugging Face

AI Mellum

  • Share
  • Facebook
  • Twitter
  • Linkedin

Prev post JetBrains AI Assistant: Smarter, More Capable, and a New Free TierJetBrains AI Assistant – Now in Visual Studio Code Next post

Subscribe to JetBrains AI Blog updates

Mellum Goes Open Source: A Purpose-Built LLM for Developers, Now on Hugging Face | The JetBrains Blog (4)

Discover more

Mellum: How We Trained a Model to Excel in Code Completion Code completion has always been a defining strength of JetBrains products. See how we trained the model behind our cloud-based completion. Anton Semenkin
How to Choose the Right LLM JetBrains AI Assistant is now multimodel! This is an exciting development, but what does it actually mean for you? It means AI Assistant is now more adaptable. It combines the strengths of different AI models, from the deep reasoning of large ones to the speed of compact ones, and even the privacy o… Irina Mariasova
AI Code Completion: Less Is More If you've been following our journey, you might've read our recent blog post on the Complete the Un-Completable: The State of AI Completion in JetBrains IDEs. From that post, you might remember this cool chart on code completions. In April, something exciting happened: our acceptance rate went up wh… Darya Rovdo
Doc-Powered AI Assistant Starting with the 2024.3 version of JetBrains IDEs, AI Assistant can not only access the context from your project but also the IDE’s own documentation. Whenever you have a question about how to accomplish a task in a JetBrains IDE, such as WebStorm, IntelliJ IDEA, or PyCharm, AI Assistant is here t… Igor Kulakov
Mellum Goes Open Source: A Purpose-Built LLM for Developers, Now on Hugging Face | The JetBrains Blog (2025)
Top Articles
Latest Posts
Recommended Articles
Article information

Author: Trent Wehner

Last Updated:

Views: 5499

Rating: 4.6 / 5 (56 voted)

Reviews: 87% of readers found this page helpful

Author information

Name: Trent Wehner

Birthday: 1993-03-14

Address: 872 Kevin Squares, New Codyville, AK 01785-0416

Phone: +18698800304764

Job: Senior Farming Developer

Hobby: Paintball, Calligraphy, Hunting, Flying disc, Lapidary, Rafting, Inline skating

Introduction: My name is Trent Wehner, I am a talented, brainy, zealous, light, funny, gleaming, attractive person who loves writing and wants to share my knowledge and understanding with you.