JetBrains has officially released the Mellum 2.1 AI model, which is designed to enhance agentic programming capabilities. The model retains the 12B mixture-of-experts architecture from its predecessor, Mellum 2, and features 2.5B active parameters. It is distributed under the Apache 2.0 license.
The development team significantly expanded the post-pretraining reinforcement learning phase, making it the core of the training process. The model was trained on datasets covering mathematics, algorithmic competitions, scientific tasks, and software engineering. JetBrains utilized an internal infrastructure that ran millions of sandboxes across thousands of environments during training.
Mellum 2.1 introduces improved agentic functionality, allowing the model to navigate codebases, edit files, and verify its own modifications. It can identify the root causes of failed tests, draft fixes, and validate the results. The integration of Multi-Token Prediction (MTP) has increased response speeds by approximately 1.6 times in single-request scenarios.
Performance benchmarks show that under high-load conditions, the inference throughput of Mellum 2.1 is nearly double that of Qwen 3.5-9B. The model also demonstrates improvements in coding, tool usage, and general knowledge compared to previous versions and competing models like Gemma 4 E4B.
The model is currently available on Hugging Face for local or private infrastructure deployment, ensuring data privacy for enterprise users. JetBrains plans to release GGUF versions compatible with llama.cpp, Ollama, and LM Studio, as well as MTP speculative decoding components for vLLM.
