The race to develop the most advanced reasoning AI is intensifying, and LG AI Research has made a bold statement with EXAONE Deep—its latest AI model designed to excel in mathematics, science, and coding. As AI moves closer to autonomous decision-making, EXAONE Deep is proving that smaller, well-optimized models can compete with industry giants.

A Step Closer to Agentic AI

The concept of Agentic AI—where artificial intelligence can think, analyze, and make decisions independently—is no longer just a theory. However, building AI with high-performance reasoning abilities is a complex challenge, and only a handful of global tech giants have made significant progress.

EXAONE Deep’s breakthrough places LG AI Research in an elite group, demonstrating that advanced reasoning AI is within reach.

How EXAONE Deep Excels in Reasoning

Unlike traditional AI models, EXAONE Deep features a dual-system architecture, mimicking the way human cognition works:
Fast, intuitive decision-making for quick responses.
Logical, methodical reasoning for complex problem-solving.

This approach enhances its ability to tackle challenging mathematical equations, scientific problems, and programming tasks—making it a valuable tool for researchers, engineers, and AI developers.

EXAONE Deep: Outperforming Bigger Models in Math

One of the most impressive aspects of EXAONE Deep is its dominance in high-difficulty math benchmarks:

32B Model Performance:
Scored 94.5 in the CSAT Mathematics section.
Achieved 90.0 in AIME 2024, matching the performance of much larger AI models.

Smaller Models Still Outperform:
7.8B model: Scored 59.6 in AIME 2025.
2.4B model: Secured 47.9, proving smaller models can deliver top-tier performance.

Why This Matters:
These results challenge the notion that bigger models are always better. EXAONE Deep’s optimized training and architecture allow it to compete with AI models five times its size—a game-changer for AI efficiency.

Strong in Science & Coding

Beyond math, EXAONE Deep has shown exceptional capabilities in science and coding:

Scientific Reasoning:
Scored 66.1 in the GPQA Diamond test (which evaluates PhD-level knowledge in physics, chemistry, and biology).

Coding Ability:
Recorded 59.5 on LiveCodeBench, outperforming other AI models in its category.

What This Means:
EXAONE Deep is not just a math model—it is an all-around problem-solving AI that can assist in scientific research, engineering, and software development.

EXAONE Deep: A Global AI Leader

EXAONE Deep has also achieved the highest MMLU (Massive Multitask Language Understanding) score among Korean AI models, solidifying LG AI Research’s position in the global AI landscape.

Recognition:

Shortly after release, EXAONE Deep was listed in Epoch AI’s Notable AI Models, an elite list of cutting-edge AI systems.
It follows the success of EXAONE 3.5, making EXAONE the only Korean AI model recognized in the past two years.

The Future of EXAONE: What’s Next?

AI is evolving at an unprecedented pace, and EXAONE Deep is leading the way.

By excelling in mathematics, science, and coding, EXAONE Deep is setting the stage for next-generation reasoning AI—one that can think, analyze, and make decisions independently.

With its record-breaking performance and efficiency, LG AI Research has firmly positioned itself as a major player in AI’s future.