LG EXAONE Deep Sets New AI Benchmarks in Maths, Science, and Coding
LG's EXAONE Deep: The Emerging AI Powerhouse Sweeping Maths, Science, and Coding Exams Off the Map
Revolutionizing Reasoning—One Test at a Time
Imagine an AI model that can solve complex maths problems, crack tricky science puzzles, and even write flawless code—all while outperforming models that are twice as large. LG AI Research has made this dream a reality with its latest innovation, EXAONE Deep, a heavyweight reasoning model that is making waves in the arena of artificial intelligence (AI).
First announced as one of the ambitious ventures of LG into high-level reasoning AI, EXAONE Deep is setting industry standards to unprecedented levels with its impressive performance in maths, science, and coding tests. But what exactly makes it so unique? Let's go into the nitty-gritty.
Small But Powerful – EXAONE Deep by LG Confronts Giants
LG AI Research has ventured into the world of AI reasoning boldly with EXAONE Deep, a model that aims to compete with the world's best. The interesting thing is that it delivers better performance while being much smaller compared to others.
Here's why EXAONE Deep is a game-changer:
Maths Mastery: EXAONE Deep 32B beat by a wide margin a model that was 20 times its size in a challenging maths benchmark. Even its younger brothers, the 7.8B and 2.4B models, won first place in all major maths benchmarks for their corresponding model sizes.
Science and Coding Excellence: In science and coding benchmarks, the 7.8B and 2.4B models excelled in all categories. EXAONE Deep's coding performance was particularly impressive in the Live CodeBench test, where it scored 59.5, outshining its capability to resolve complex programming issues.
MMLU Dominance: The 32B model achieved a score of 83.0 on the Massive Multitask Language Understanding (MMLU) benchmark, the highest of any local Korean model.
Benchmark Breakdown: EXAONE Deep's Stellar Performance
LG AI Research not only constructed EXAONE Deep—it trained it to dominate real-world usage. Here a more detailed analysis of how it performed on various tests:
Mathematics: Precision and Power
In mathematics, EXAONE Deep illustrated its problem-solving and logical capabilities:
The 32B model scored 94.5 on a general mathematics proficiency test and 90.0 on the American Invitational Mathematics Examination (AIME) 2024.
In the AIME 2025, it kept pace with the performance of DeepSeek-R1, a gigantic 671B model, to show that size isn't everything when it comes to reasoning efficiency.
The 7.8B model performed at 94.8 on the MATH-500 benchmark and 59.6 on AIME 2025, while the 2.4B model earned scores of 92.3 and 47.9 on the same tests—making them class leaders in their weight divisions.
Science and Coding: Shattering Expectations
EXAONE Deep isn't just a math wizard—its also a science and computer programming genius:
The 32B scored 66.1 on the GPQA Diamond test, a measure of advanced problem-solving skills in doctoral physics, chemistry, and biology.
It achieved 59.5 on Live CodeBench, a coding skills test, indicating its potential for practical software development and automation applications.
Even the smaller 7.8B and 2.4B models led in their respective categories on the GPQA Diamond and Live CodeBench tests, solidifying their supremacy.
General Knowledge: Increased Overall Intelligence
Other than technical exams, EXAONE Deep also acts as a decent general knowledge model:
It scored 83.0 on the MMLU test, the highest-scoring native Korean model and cementing its reputation as a general knowledge comprehended with versatile capabilities.
Why It Matters: LG's Global Recognition
EXAONE Deep's performance has not escaped attention. It was soon after its release listed under the 'Notable AI Models' category by US-based nonprofit Epoch AI, along with its predecessor EXAONE 3.5. This renders LG the sole Korean organization to be recognized on this elite list twice within two years—a major milestone for Korea's AI industry.
What This Means for AI’s Future
LG's EXAONE Deep is not only a super-performance model—it's the future of AI reasoning. Being able to perform better than bigger models shows that the future of AI innovation is no longer about how big but how efficient and intelligent.
Furthermore, being adept at coding and science makes it such a useful tool in automating difficult programming tasks, aiding scientific research, and even education and STEM.
SOURCE BY:- AI NEWS
The Road Ahead for AI Reasoning
As LG AI Research continues to break new ground with EXAONE Deep, the question is: How far can AI reasoning go? With its unprecedented performance and efficiency, EXAONE Deep is leading the way for more efficient yet powerful models that can address real-world challenges with human-like accuracy.
At SkillBloomer, we’ll keep you updated on the latest breakthroughs in AI, coding, and science. Stay tuned for more insights into the future of artificial intelligence and how it’s shaping our world.
