EXAONE Deep
Reasoning-focused language models from LG AI Research
EXAONE Deep is a series of reasoning-enhanced language models released by LG AI Research in 2025. The family includes 2.4B, 7.8B and 32B parameter variants trained on data rich in step-by-step thought processes. It aims to improve performance in mathematics and coding tasks while remaining openly available for research use.
- Released by LG AI Research in March 2025
- Includes 2.4B, 7.8B and 32B parameter models
- Trained on datasets with long reasoning chains
- Outperforms similar-sized models in math and code
- Open weights available for research use
What it is
EXAONE Deep is a family of large language models designed to excel at complex reasoning tasks. Developed by LG AI Research, the series was introduced in March 2025. It comprises three sizes: 2.4 billion, 7.8 billion and 32 billion parameters. The models focus on improving capabilities in mathematics and computer programming by learning from detailed chains of thought. They are made available as open weights for the research community.
How it came about
The project was led by a team including Kyunghoon Bae and Jinsik Lee among others. The initial preprint appeared on arXiv in March 2025. A revised version was published in January 2026. The work presents the full series and its evaluation results, marking a shift towards reasoning-specific training data for the EXAONE line.
Inside the model
The models use a standard decoder-only transformer architecture. Training relied heavily on a specialised dataset containing long sequences of reasoning steps. This approach teaches the system to break down problems rather than just predicting the next word. The team trained three distinct model sizes to test scaling effects. The largest variant contains 32 billion parameters.
What it can do
Evaluation shows that the smaller 2.4B and 7.8B models beat other systems of similar size on math and coding tests. The 32B model performs competitively against leading open-weight alternatives. The primary strength lies in handling tasks that require multi-step logical deduction. Performance gains are attributed to the thought-process-heavy training data.
How to use it
All models in the EXAONE Deep series are openly available. Researchers can download the weights directly for study and experimentation. The release is intended for academic and investigative use rather than commercial deployment without further licensing. The files are hosted on public repositories linked from the project paper.
Questions people ask about EXAONE Deep
What makes EXAONE Deep different from other models?
It is trained primarily on a dataset filled with long streams of thought processes. This focuses the model on reasoning steps rather than just pattern matching, leading to better performance in math and coding tasks.
How large are the EXAONE Deep models?
The series includes three sizes. The smallest has 2.4 billion parameters, the medium has 7.8 billion, and the largest contains 32 billion parameters. Each is designed to offer competitive performance within its size class.
Can I use EXAONE Deep for commercial products?
The sources state the models are openly available for research purposes. They do not specify a commercial licence. Users should check the specific licence terms associated with the downloaded weights for any restrictions on commercial use.
Good reading and sources
There is no English Wikipedia article on EXAONE Deep yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2503.12524. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.