EXAONE 4.0 1.2B is a compact, on-device language model developed by LG AI Research. Released as part of the EXAONE 4.0 series in July 2025, it is designed for high-efficiency performance on consumer hardware like smartphones and home appliances. The model features a hybrid architecture that integrates two distinct operational states: a non-reasoning mode for rapid, knowledge-based tasks and a reasoning mode for complex problem-solving.
In its non-reasoning configuration, the model is optimized for general natural language processing, instruction following, and world knowledge retrieval. This mode provides low-latency responses for standard interactions, such as summarization and general Q&A, across multiple languages including Korean, English, and Spanish. It is engineered to deliver concise and accurate outputs by leveraging the model's foundational pre-training without the additional computational overhead of step-by-step reasoning blocks.
The model supports a context window of 65,536 tokens (64K) and utilizes a "QK-Reorder-Norm" architecture, which applies normalization directly to attention and MLP outputs to enhance stability and performance. Despite its small size, it supports agentic tool use and function calling, allowing it to interact with external tools in various software environments.
Developed as a successor to EXAONE 3.5, the 1.2B variant achieved improved benchmark scores in mathematics, coding, and science while being significantly smaller than its 2.4B predecessor. It was trained through a multi-stage process involving supervised fine-tuning and preference learning, focusing on maintaining objectivity and fairness in its multilingual capabilities.