Jetbrains Open Source: Developer-centric Language Model for Code-related Tasks

Jetbrains has officially opened source Mellera dedicated 4 billion parameter language model tailored for software development tasks. Mellum starts from scratch, reflecting Jetbrains’ engineering-first approach, providing a domain-specific model that trains practical uses across password libraries and programming environments. Jetbrains, released on the surface under the Apache 2.0 license, offers invitations to the wider community of research and developer to experiment, adapt and enhance Mellum’s capabilities.

Focus Model for Code Understanding

Unlike general LLM, Mellum is classified as a “focus model” by Jet Brain, which they use to describe models with narrow but profound specialization. Mellum is optimized specifically for programming-related tasks such as autocomplete, fill and structural understanding of source code. This focused design avoids the overhead of broader language modeling and enables the model to behave effectively in an IDE-like environment.

The model supports a wide variety of languages ​​including Java, Kotlin, Python, GO, PHP, C, C, C++, C#, JavaScript, Typescript, Typescript, CSS, HTML, Ruby and Ruby, promoting the multi-faceted nature of modern development teams.

Model architecture and training pipeline

Mellum 4.2 trillion tokens Taken from a rich source of code like Stack, Starcoder, commitpack, and English Wikipedia. It has an 8K token context window and uses BF16 Mixing Accuracy High-throughput cluster of 256 NVIDIA H200 GPUs connected via Infiniband.

The training process is approximately 20 days and scalable model development is carried out using modern infrastructure. The build and training process is designed with reproducibility and deployment flexibility, making Mellum available in two cloud inference settings (e.g., VLLM) and on-premises environments (e.g., Llama.cpp, Ollama).

Benchmarks and evaluations

Jetbrains evaluates Mehlham, a range of benchmark tests that reflect its primary use cases (coded fill and finished). The performance of this model shows strong consistency with the design objectives:

• Review of V1.1 (8K context):
  • Python EM: 27.97%
  • Java EM: 31.08%
• Safim (Syntax-aware padding):
• Humane filling:
  • Single line: 66.21%
  • Multiple lines: 38.52%
  • Random span: 29.70%

These results reflect Mellum’s specialization in structured code understanding, especially when partial or interrupted code is involved, which is common in real-world development workflows.

Reasons for opening up procurement

Jetbrains’ decision to use Mellum as open source was based on several practical motivations:

• transparency: Review of training data and building decisions.
• Reusable: Supports integration in custom development environments and research experiments.
• Community Cooperation: Promote external developers’ contribution to improving model behavior.
• Teaching value: Provide hands-on products for educators and students to understand how to build and apply domain-specific LLMs.

The release includes Basic Model (Mellum-4b bas) and a Fine-tuning variants For Python (Mellum-4b-Sft-Python).

Impact on developer tools

Compact optimization for source code, the availability of performance models opens up new opportunities for IDE space and other regions. Jetbrains envisions Mellum as part of a broader strategy involving multiple focus models, each optimized for specific programming tasks, such as DIFF generation or code review help. This approach is associated with growing demand for deployable, cost-effective and context-aware AI tools that can increase developer productivity without introducing opaque or super-large general models.

in conclusion

Mellum represents a deliberate shift towards a smaller professional language model that prioritizes utility, transparency, and efficiency. By publicly providing the model, JetBrains provides a high-quality foundation for building the next generation of AI-assisted developer tools. Its architecture, training methodology and benchmark performance indicate a practical step in the ever-evolving space of LLMS tailored to software engineering.

The release includes Basic Model (Mellum-4b bas) and a Fine-tuning variants For Python (Mellum-4b-Sft-Python). Also, don’t forget to follow us twitter And join us Telegram Channel and LinkedIn GrOUP. Don’t forget to join us 90K+ ml reddit.

🔥 [Register Now] Minicon Agesic AI Virtual Conference: Free Registration + Certificate of Attendance + 4-hour Short Event (May 21, 9am-1pm) + Hands-On the Workshop