JetBrains Releases Open-Source Mellum2.1 Coding AI Model Optimized for Agents
JetBrains has officially released Mellum2.1, an open-source (Apache 2.0) 12B Mixture-of-Experts AI model with 2.5B active parameters designed for agentic software engineering. The model features major post-training reinforcement learning upgrades and delivers high-load inference throughput nearly double that of Qwen3.5-9B. Mellum2.1 enables software developers and enterprises to run fast, competent open-weights coding agents locally or on private infrastructure without risking data privacy. Its enhanced agentic capabilities allow it to autonomously diagnose test failures, draft code fixes, and verify edits across full codebases. Leveraging Multi-Token Prediction (MTP), Mellum2.1 boosts single-request speeds by 1.6x while maintaining high throughput under heavy concurrent workloads. JetBrains trained the model using millions of sandboxed environment runs for reinforcement learning and announced upcoming support for GGUF formats and vLLM MTP speculative decoding.
## BACKGROUND
Mixture-of-Experts (MoE) architectures route prompts to specialized sub-networks within the model, activating only a small fraction of overall parameters to substantially reduce hardware demands during inference. Meanwhile, Multi-Token Prediction (MTP) is an inference optimization paradigm where models learn to predict multiple subsequent tokens simultaneously rather than generating output one token at a time.