LG AI Research recently open-sourced its acclaimed EXAONE Deep reasoning AI model, touted as a "globally renowned" achievement! Its biggest highlight is its reasoning capabilities. Simply put, it can act like a clever detective, independently formulating hypotheses and testing them to make autonomous decisions. This marks AI's entry into a new era of Agentic AI, suggesting that AI might truly start "thinking for itself" in the future!
LG AI Research proudly announces that EXAONE Deep is the first self-developed reasoning AI model in South Korea to achieve this milestone.
32 Billion Parameters
The EXAONE Deep model, with its 32 billion parameters (32B), is surprisingly only about 5% the size of some "giant" models!
Even more astonishing is this "lightweight" model's remarkable ability to solve complex problems. Performance evaluations published on the global open-source AI platform Hugging Face show that EXAONE Deep-32B excels in logical reasoning and problem-solving. Notably, it achieved a top score of 94.5 out of 100 in the South Korean College Scholastic Ability Test (CSAT) mathematics section and received the highest grade in all elective subjects! This is comparable to a top-tier STEM student. Furthermore, it scored 95.7 on the MATH-500 benchmark, which assesses mathematical problem-solving skills.
A Research Favorite: Mastering Math, Science, and Programming
EXAONE Deep-32B's "top student" status doesn't end there. It also performed exceptionally well in benchmarks assessing doctoral-level scientific reasoning in physics, chemistry, and biology. It scored 66.1 on the GPQA diamond test, which evaluates scientific problem-solving abilities, surpassing other reasoning AI models of comparable size. It even achieved a commendable score of 59.5 on LiveCodeBench, a coding ability assessment.
It's no wonder LG AI Research confidently states that EXAONE Deep will shine in industry-specific professional fields, as well as in research and education in areas like physics and chemistry. With such impressive capabilities, EXAONE Deep-32B has been listed on the "Notable AI Models" list compiled by Epoch AI, a US-based non-profit AI research organization.
Lightweight and On-Device Models Equally Impressive
Besides the powerful 32B version, LG AI Research also open-sourced a "lightweight model, EXAONE Deep-7.8B" and an "on-device model, EXAONE Deep-2.4B". These models are like "mini versions" of the 32B model; despite their significantly smaller size, their performance remains noteworthy. The lightweight model, at only 24% the size of the 32B model, maintains a remarkable 95% of its performance. The even smaller on-device model, at just 7.5% the size, boasts 86% of the performance.
This is the perfect example of "the best things come in small packages," offering both economy and practicality. LG AI Research officials also revealed that they had previously announced plans to open-source the DeepSeek R1-level model in February, demonstrating their commitment to their promises. They emphasize that significantly reducing model size while maintaining performance is key to LG's AI technology.
It's worth mentioning that the "mini" on-device model offers unique security advantages. Because data can be processed securely on the device without connecting to external servers, it better protects user security and personal information. LG AI Research anticipates that this type of on-device AI model will rapidly gain popularity across various industries, including smartphones, automobiles, and robotics. They also plan to collaborate with affiliated companies like LG Electronics and LG U+ to further enhance these models and aim for a leading position in the on-device AI market.
Project Link: https://top.aibase.com/tool/exaone-deep
