LG AI Research recently announced the open source of its highly anticipated EXAONE Deep Inference AI model, which is well-known for its outstanding reasoning capabilities. EXAONE Deep can make independent decisions like a clever detective, making assumptions and conducting inference verification. This breakthrough marks that AI technology is entering a new era of "active AI", and AI may have stronger independent thinking ability in the future.
LG AI Research proudly stated that EXAONE Deep is the first self-developed inference AI model in South Korea to achieve this milestone. This model not only made a major breakthrough in technology, but also attracted widespread attention worldwide.
EXAONE Deep has 32 billion parameters (32B), and although this parameter is only 5% of some "bigma" models, it has shown amazing strength in solving complex problems. Performance evaluation results released on Hugging Face, a global open source AI platform, show that EXAONE Deep-32B has performed outstandingly in logical reasoning and problem-solving capabilities. Especially in the field of mathematics in the Korean College Entrance Examination (CSAT), it actually scored the highest score of 94.5 and reached the highest level among all elective subjects. This score is comparable to a master of math, physics and chemistry. In addition, it also scored 95.7 points in the MATH-500 index for evaluating mathematical problem-solving ability.
EXAONE Deep-32B not only performs well in the field of mathematics, but also becomes a "new favorite" in the scientific research community. It also performs well in physical, chemical and biological indicators that evaluate the ability of scientific reasoning at doctoral level. In the GPQA diamond test that evaluates scientific problem-solving ability, it scored 66.1 points, surpassing other inference AI models of the same parameter scale. Even in LiveCodeBench, which evaluates code capabilities, he also achieved a good score of 59.5.
LG AI Research said with confidence that EXAONE Deep will shine in the professional fields required by industry and in the fields of scientific research and education such as physics and chemistry. With such hard-core strength, EXAONE Deep-32B also successfully listed on the "Notable AI Models" list selected by Epoch AI, a non-profit AI research organization in the United States.
In addition to the powerful 32B version, LG AI Research also open sourced the "lightweight model EXAONE Deep-7.8B" and the "end-side model EXAONE Deep-2.4B". These two models are like the "mini version" of 32B. Although the size is greatly reduced, the performance is still remarkable. The lightweight model maintains up to 95% performance when it is only 24% of the size of the 32B model. The smaller end-side model also has 86% performance when it is only 7.5%.
This is simply the best interpretation of "all the concentrates are the essence", both economical and practical. Officials from LG AI Research also revealed that as early as February this year, they had previewed the upcoming plan to open source the DeepSeek R1-level model, which now seems to be "doing what you say." They stressed that maintaining performance while significantly reducing model size is the key to LG AI technology.
It is worth mentioning that this "mini" end-side model has unique advantages in terms of safety. Since data can be processed securely within the device without connecting to an external server, users' security and personal information can be better protected. LG AI Research expects that this end-side AI model will be rapidly popularized in various industries such as smartphones, cars and robots. They also plan to work with affiliates such as LG Electronics and LG U+ to further enhance these models in order to take the lead in the end-side AI market.
Project entrance: https://top.aibase.com/tool/exaone-deep