LG AI Research has introduced new reasoning models (LRMs) named EXAONE Deep. Available in three sizes (2.4B, 7.8B and 32B parameters), the models have been optimized for math, science and programming, as is usual with LRMs. According to the company, the largest model, EXAONE Deep 32B, can keep up with DeepSeek R1 in some benchmarks, but is beaten by the significantly larger model in others. The more compact versions should beat OpenAI’s outdated o1-mini and distill variants such as DeepSeek-R1-Distill-Qwen. Unlike the DeepSeek models, LG’s models are covered by the EXAONE AI Model License Agreement 1.1 – NC, which excludes commercial use. All EXAONE Deep models are available for research purposes on Hugging Face.
