LGAIResearch, yes, that LG that develops consumer electronics, launched EXAONEDeep, a high-performance reasoning artificial intelligence that, despite a relatively small number of parameters, has demonstrated extraordinary capabilities in mathematical logic, scientific concepts, and programming challenges.
The performance metrics of the flagship 32B model are comparable to larger models such as GPT-4o and DeepSeekR1. In comparison, the 7.8B and 2.4B variants set new benchmarks in the lightweight and on-device AI categories.
The EXAONEDeep32B model scored 94.5 points in the math section of CSAT2025 and 90.0 points in AIME2024, outperforming other competing models while requiring only 5% of the computing resources of large alternative models such as DeepSeek-R1 (671B).
In scientific reasoning, it achieved a score of 66.1 on the GPQA Diamond test, which assesses PhD-level problem-solving skills in physics, chemistry and biology. The model scored 83.0 points in MMLU, the highest score among models developed in Korea.
Of particular note is the performance of the smaller variants: the 7.8B model scored 94.8 points in MATH-500 and 59.6 points in AIME2025, while the 2.4B model scored 92.3 points in MATH-500 and 47.9 points in AIME2024. These results place EXAONEDeep's smaller models at the top of their categories on all major benchmarks, demonstrating the huge potential for deploying EXAONEDeep in resource-constrained environments.
EXAONEDeep has up to 32 billion parameters and performs well in single-GPU deployments. Interestingly, these models can run on a range of discrete GPUs, laptop GPUs, and some edge systems that do not have large-scale computing capabilities.