Discover Mellum, JetBrains’ pioneering open-source LLM specifically designed for coding needs and development tasks.
Mellum-4b-base, JetBrains’ first open-source large language model optimized for code, leverages over 4 trillion tokens and an 8192-token context window across multiple programming languages. Built on a LLaMA-style architecture with 4 billion parameters, this efficient model (using Automatic Mixed Precision, AMP, bf16) is ideal for cloud inference (e.g., via vLLM) and local deployment (e.g., llama.cpp or Ollama), particularly for code completion. Designed for integration into developer tools, AI-powered coding assistants, and research on code understanding and generation, Mellum also serves educational applications and fine-tuning experiments. The initial release includes a base model and Python SFT models, with further language support planned. While the base model isn’t pre-tuned, it’s fully compatible with supervised fine-tuning (SFT) and reinforcement learning (RL).
