Skip to content
Search

EXAONE Deep

lastmod 2025-03-19created 2025-03-191 edits

A new reasoning AI model developed by LG AI Research

About

  • LG AI Research releases EXAONE Deep, a new reasoning AI model it developed
  • A high-performance reasoning model essential to the transition to the agentic AI era
  • Demonstrated excellent reasoning ability in math, science, and coding
  • Key features:
    • Math: outperforms competing models on high-difficulty math benchmarks (achieves equivalent performance with a smaller model size)
    • Science & coding: ranked 1st on major benchmarks (7.8B and 2.4B models)
    • MMLU: best performance among Korean models (32B model)
  • Performance validated by being listed among Epoch AI's notable AI models (the second EXAONE model after EXAONE 3.5)

Outstanding performance in math

  • All EXAONE Deep models (32B, 7.8B, 2.4B) achieved top scores in the math section of the 2025 College Scholastic Ability Test (CSAT)
  • EXAONE Deep 32B:
    • Scored 94.5 on CSAT math and 90.0 on AIME 2024
    • Performance on par with the DeepSeek-R1 (671B) model on AIME 2025
    • Showed especially strong performance on AIME, a high-difficulty benchmark, demonstrating training efficiency and cost efficiency
  • EXAONE Deep 7.8B & 2.4B:
    • Ranked 1st on major benchmarks in the lightweight model and on-device model categories respectively
    • 7.8B model: 94.8 on MATH-500, 59.6 on AIME 2025
    • 2.4B model: 92.3 on MATH-500, 47.9 on AIME 2024
  • Major math benchmarks:
    • CSAT (College Scholastic Ability Test)
    • AIME (American Invitational Mathematics Examination)
    • MATH-500

Outstanding expertise in science and coding

  • Also demonstrated superior performance over competing models in science and coding
  • EXAONE Deep 32B:
    • 66.1 on the GPQA Diamond test (evaluates PhD-level science problem-solving ability)
    • 59.5 on LiveCodeBench (evaluates coding ability), surpassing reasoning AI models of similar size
    • Suggests high applicability in fields requiring specialized knowledge
  • EXAONE Deep 7.8B & 2.4B:
    • Ranked 1st on GPQA Diamond and LiveCodeBench
    • Following the EXAONE 3.5 2.4B model, established as a global leading model in the lightweight/on-device model field
  • Major science and coding benchmarks:
    • GPQA Diamond
    • LiveCodeBench

Documentation

[2503.12524] EXAONE Deep: Reasoning Enhanced Language Models
https://arxiv.org/abs/2503.12524

See also

Favorite site

author Zer0 sitemap