JetBrains Releases Mellum2, a 12B MoE Model for Low-Latency Text and Code

JetBrains has unveiled Mellum2, a Mixture-of-Experts (MoE) model featuring 12 billion parameters, developed from the ground up for natural language and code. This model activates just 2.5 billion parameters per token, facilitating efficient, low-latency inference. It is released under the Apache 2.0 license and can be found on Hugging Face. Mellum2 is aimed at latency-sensitive tasks such as routing, retrieval-augmented generation (RAG), summarization, sub-agents, and private deployments, boasting over 2x faster inference compared to similar-sized open models. Focused solely on text and code, it avoids multimodal tasks to maintain efficiency in software engineering. A comprehensive technical report on its architecture, training, benchmarks, and evaluation is accessible on arXiv.