EXAONE 4.0 1.2B is a compact, open-weight language model developed by LG AI Research and released in July 2025. It serves as the on-device variant of the EXAONE 4.0 series, optimized for high performance on edge devices and in resource-constrained environments. The model is characterized by its "hybrid" architecture, which allows it to toggle between a standard non-reasoning mode for fast responses and a specialized Reasoning mode for complex logical tasks.
In its Reasoning mode, the model utilizes a step-by-step thinking process to excel in domains such as mathematics, coding, and scientific deduction. This variant is specifically fine-tuned using reasoning reinforcement learning and preference learning to generate structured thought blocks before providing a final answer. It supports a multilingual suite including Korean, English, and Spanish, and incorporates native support for agentic capabilities like function calling and the Model Context Protocol (MCP).
Architecturally, the model features approximately 1.2 billion parameters and supports a context window of up to 64,000 tokens. It employs technical refinements such as QK-Reorder-Norm, which reorders LayerNorm positions and applies RMS normalization to query and key projections to improve training stability and performance on downstream benchmarks. Despite its small footprint, it is designed to achieve results comparable to larger models in specialized reasoning evaluations.