JetBrains has released Mellum2, a 12 billion-parameter Mixture-of-Experts model optimized for natural language and coding tasks. With efficient parameter activation and over 2x faster inference compared to similar models, Mellum2 is positioned for high-throughput AI applications.
Mellum2 is a newly released 12 billion parameter Mixture-of-Experts (MoE) model developed by JetBrains. It is specifically designed to handle a variety of natural language and coding tasks while optimizing for low-latency inference, making it suitable for production environments.
The model's architecture allows it to activate only 2.5 billion parameters for each token, greatly enhancing its efficiency compared to traditional models. Mellum2 has been benchmarked against similar-sized models and delivers competitive performance while achieving more than twice the inference speed.
Mellum2 is applicable for a range of functions including routing, retrieval-augmented generation (RAG), summarization, and even coding features. It supports latency-sensitive tasks which are crucial for real-time applications, demonstrating its adaptability in various AI systems.
Mellum2 is available for download under the Apache 2.0 license from Hugging Face. Full architectural details, training setups, and evaluation methodologies can be accessed through the accompanying technical report.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
JetBrains has released Mellum2, a 12 billion-parameter Mixture-of-Experts model optimized for natural language and coding tasks. With efficient parameter activation and over 2x faster inference compared to similar models, Mellum2 is positioned for high-throughput AI applications.