by JetBrains
Explore Mellum, a family of fast language models, including a next-generation model for ultra-low-latency and high-performance inference.
The open-source LLM by JetBrains, optimized for real-world development workflows where latency and performance matter most.
Mellum understands code, context, and intent, expanding beyond pure code completion to support both natural language and programming tasks.
A mixture-of-experts (MoE) architecture delivers ultra-low-latency inference and high throughput, often twice as fast as similar-sized models, while bringing MoE capabilities to a much smaller model class.
Mellum achieves strong coding quality while halving inference costs, thanks to fewer active parameters per request and efficient compute utilization.
Trained on transparent data and aligned for consistency, Mellum can be fine-tuned and deployed locally or on the cloud, giving you full control over performance, privacy, and infrastructure.
Built for AI/ML engineers and researchers, Mellum powers efficient, real-world AI workloads, from code completion to broader language and programming tasks, and supports both local and cloud deployment.
A 12B-parameter open-source mixture-of-experts model, designed for real-time workflows, combining strong coding and language capabilities with exceptional efficiency.
An open-source coding model built for broad code understanding and completion across multiple languages.
Analyze incoming prompts and select the right model for each task for fast, intelligent routing between models based on use case and requirements.
Retrieve relevant information and use Mellum to quickly summarize and generate responses, keeping question-answering systems fast and responsive.
Break down agent pipelines into steps, like context gathering, planning, and validation, and use Mellum for fast, specialized tasks instead of relying on a single large model.
Keep code and data fully under your control with local or self-hosted deployment for private and sovereign AI use cases.
We built Mellum because not every task requires the largest or most complex models. By focusing on performance, latency, and cost, we created a model designed for developers and teams moving from experimentation to production.
JetBrains Mellum is a family of fast language models designed for efficient, real-world AI workloads. What started as a code-focused model has evolved to support both programming and natural language tasks, enabling fast, cost-efficient inference across a wide range of use cases.
The latest Mellum model introduces a new mixture-of-experts architecture optimized for performance and efficiency. It expands beyond code completion to support broader language and programming tasks, and is designed for low-latency, real-time use cases and production deployment.
Large models are powerful, but expensive and run more slowly at scale.
Mellum is designed for scenarios where:
The latest generation of Mellum is pre-trained from scratch using a mix of natural language and code data, with a focus on coding and mathematical domains. We then further improve model quality through post-training to make it competitive with other small open-source LLMs.
We pre-train Mellum2 on more than 10 trillion tokens of permissively licensed code, web text, and math. Pre-training consists of three stages: first, we train the model on ~6T tokens of mostly web data; then we follow that with 2.8T tokens of high-quality data with a strong focus on coding. Finally, we train on 1.7T tokens of the highest-quality data, where code accounts for more than 50% of the tokens.
This is followed by large-scale post-training: first, supervised fine-tuning (SFT) on 50B tokens of instruction-following data, then reinforcement learning with verifiable rewards (RLVR) across diverse coding, math, and reasoning environments.
Together, these stages of broad pre-training, progressive concentration on code and math, and targeted post-training with SFT and RLVR ensure that Mellum2 is compact and a strong contender among small open-source LLMs.
If you want to learn more about how Mellum1 was trained, read the blog post.
Mellum uses a mixture-of-experts design with only a small subset of active parameters per request, reducing compute requirements while maintaining high output quality. On a GPU, VRAM has to hold both the model weights and the KV-cache, which grows with every parallel request and every token of context. We deliberately designed Mellum with a compact KV-cache footprint per request, so a single GPU can serve many more concurrent users without running out of memory. This enables highly efficient deployment across H100/H200 GPUs as well as newer B200 and B300 hardware – even on a single GPU.
Mellum2: General chat-interface usage in English and all major programming languages.
Mellum1: Java, Kotlin, Python, Go, PHP, C, C++, C#, JavaScript, TypeScript, CSS, HTML, Rust, and Ruby.
Yes. Mellum is released under the Apache 2.0 license, with open weights available on Hugging Face.
You can also try Mellum out of the box:
Run Mellum locally or on the cloud. It’s built for real-time systems, efficient inference, and modern AI development.