Source-linked AI summary
EXAONE 4.5 Technical Report
Eunbi Choi, Kibong Choi, Sehyun Chun, Seokhee Hong, Junwon Hwang, Hyojin Jeon, Ahra Jo, Hyunjik Jo, Yeonsik Jo, Joonkee Kim, Seonghwan Kim, Soyeon Kim, Sunkyoung Kim, Yireun Kim, Yongil Kim, Changhun Lee, Haeju Lee, Jinsik Lee, Kyungmin Lee, Sangha Park, Kwangrok Ryoo, Minju Seo, Sejong Yang, Heuiyeen Yeen, Hwan Chang, Stanley Jungkyu Choi, Yejin Choi, Kyubeen Han, Joonwon Jang, Kijeong Jeon, Geunyeong Jeong, Gerrard Jeongwon Jo, Jiyeon Jung, Daeseong Kim, Dohoon Kim, Dohyun Kim, Hyunseo Kim, Minu Kim, Myoungshin Kim, Youchul Kim, Byungoh Ko, Christopher Lee, Edward Hwayoung Lee, Honglak Lee, Jiyoung Lee, Sangeun Lee, Seungwon Lim, Woohyung Lim, Jueun Mun, Jaewoo Park, Jimin Park, Jinho Park, Yongmin Park, Wooseok Seo, Yongwoo Song, Sihyuk Yi, Kyungjae Yoo, Sangyeon Yoon
TL;DR
EXAONE 4.5 addresses the challenge of preserving rich visual information while extending multimodal reasoning to long contexts. It integrates a billion-parameter-scale vision encoder with the EXAONE 4.0 language model and achieves competitive performance across vision and language tasks, including outperforming larger or closed-weight models in complex domains.
Problem
Vision-language models must process many visual tokens as image resolution increases, while aggressive token reduction discards critical spatial and semantic information.
Method
EXAONE 4.5 integrates a custom-built 1.2B-parameter vision encoder with the EXAONE 4.0 32B language model and uses multi-stage multimodal pretraining with context extension to 256K tokens.
Results
EXAONE 4.5 exhibits highly competitive, state-of-the-art performance across diverse vision and language tasks, frequently outperforming substantially larger or closed-weight models in complex domains.
Takeaways & Limitations
As an open-weight model, EXAONE 4.5 is intended to support demanding industrial problem solving, community-driven research, and future Vision-Language-Action models.
Takeaways & Limitations
Multilingual evaluations cover only Korean, German, Spanish, Japanese, and Vietnamese.
Abstract
from arXiv · showhide
This technical report introduces EXAONE 4.5, the first open-weight vision language model released by LG AI Research. EXAONE 4.5 is architected by integrating a dedicated visual encoder into the existing EXAONE 4.0 framework, enabling native multimodal pretraining over both visual and textual modalities. The model is trained on large-scale data with careful curation, particularly emphasizing document-centric corpora that align with LG's strategic application domains. This targeted data design enables substantial performance gains in document understanding and related tasks, while also delivering broad improvements across general language capabilities. EXAONE 4.5 extends context length up to 256K tokens, facilitating long-context reasoning and enterprise-scale use cases. Comparative evaluations demonstrate that EXAONE 4.5 achieves competitive performance in general benchmarks while outperforming state-of-the-art models of similar scale in document understanding and Korean contextual reasoning. As part of LG's ongoing effort toward practical industrial deployment, EXAONE 4.5 is designed to be continuously extended with additional domains and application scenarios to advance AI for a better life.
1 Introduction
EXAONE 4.5 extends the EXAONE foundation-model series with visual comprehension as LG’s first open-weight VLM. Its architecture and training pipeline combine a billion-parameter-scale vision encoder, multimodal alignment, document- and OCR-focused data, and stable 256K-token context extension.
- Model Motivation: EXAONE 4.5 is LG’s first open-weight Vision Language Model, extending the EXAONE 4.0 hybrid LLM with visual comprehension.EXAONE 4.0 provided separate non-reasoning and reasoning modes; EXAONE 4.5 expands this foundation beyond text.
- Architecture: A 1.2B-parameter vision encoder is integrated with the EXAONE 4.0 32B language model to process text and image inputs.The encoder was trained from scratch because existing vision encoders did not meet the required scalability and efficiency.
- Architecture: The vision encoder uses 2D RoPE and GQA, while the language model retains 1D RoPE to capture image structure and preserve textual compatibility.GQA is also used to reduce attention complexity and improve hardware utilization in the vision encoder.
- Training Pipeline: Multimodal pretraining progresses from foundational visual-text alignment to perceptual and knowledge refinement using document, OCR, grounding, mathematics, and STEM data.The curriculum includes general image-text pairs, interleaved image-text documents, and task-oriented images such as charts, diagrams, and document-parsing examples.
3 Evaluation
EXAONE 4.5 is evaluated across broad vision and language benchmark categories, including reasoning, document understanding, Korean, long-context, and tool-use tasks. Results show balanced vision performance and particularly strong language performance in coding, mathematical reasoning, tool use, and instruction following.
- Vision Evaluation: Vision evaluation covers STEM and puzzle reasoning, document understanding, general perception, and Korean benchmarks.The document category includes AI2D, CHARTQAPRO, CHARXIV (RQ), OCRBENCH V2, and OMNIDOCBENCH V1.5.
- Language Evaluation: Language evaluation covers reasoning, agentic tool use, instruction following, long-context understanding, Korean, and multilinguality.The benchmark suite includes AIME 2026, GPQA-DIAMOND, LIVECODEBENCH V6, τ2-BENCH, IFBENCH, IFEVAL, and AA-LCR.
- Vision Results: 75.2 vs. 74.6 on MATHVISION and 79.1 vs. 74.8 on WE-MATH versus Qwen3-VL-235B demonstrate strong multimodal mathematical reasoning despite the baseline’s larger size.EXAONE 4.5 also exceeds GPT-5 mini on MMMU-PRO (68.6 vs. 67.3) and MATHVISION (75.2 vs. 71.9).
- Vision Results: 71.7 vs. 66.1 on CHARXIV (RQ), 89.0 vs. 88.2 on AI2D, and 81.2 vs. 77.0 on OMNIDOCBENCH show strong document understanding against larger or closed-weight baselines.EXAONE 4.5 also outperforms Qwen3-VL-235B on BLINK (68.7 vs. 67.1), indicating stable general vision performance.
- Language Results: EXAONE 4.5 achieves the best score on LIVECODEBENCH V6 and ranks second overall on AIME 2026, indicating strength in coding and mathematical reasoning.It ranks first among all compared baselines on AIME 2026 except Qwen3.5-27B.
- Language Results: 72.0 versus 57.0 on weighted-average τ2-BENCH and higher IFBENCH and IFEVAL scores show strong agentic tool-use and instruction-following performance.The instruction-following scores are 62.6 vs. 59.2 on IFBENCH and 89.6 vs. 88.2 on IFEVAL against Qwen3-VL-235B-A22B.
4 Limitations
EXAONE 4.5 has limitations involving inappropriate, biased, statistically driven, outdated, or contradictory outputs, as well as restricted multilingual evaluation coverage.
- Multilingual benchmarks evaluate only Korean, German, Spanish, Japanese, and Vietnamese among non-English supported languages.
- Problematic personal, harmful, or biased content may remain in training data despite efforts to exclude it, potentially producing undesirable responses.
- EXAONE 4.5 may generate inappropriate answers containing personal, harmful, biased, or other inappropriate information.
- Generated responses rely heavily on training-data statistics, which can produce semantically or syntactically incorrect sentences.
- Because the models do not reflect the latest information, their responses may be false or contradictory.
5 Deployment
The report directs users to consult the model license information before using EXAONE 4.5 legally.
- License information for using the EXAONE 4.5 models is provided in Appendix Section B.
- Understanding the license information is essential for the legal utilization of the language model.
- Deployment therefore requires attention to the applicable model license before use.
6 Conclusion
EXAONE 4.5 combines a 1.2B-parameter vision encoder with the EXAONE 4.0 32B language model and uses architectural and training innovations for multimodal reasoning. Broad evaluations report highly competitive performance across vision and language tasks, including complex mathematical reasoning, document parsing, and agentic tool use, while the open-weight release supports future VLA development and community research.
- Architecture and training: EXAONE 4.5 integrates a custom-built 1.2B-parameter vision encoder with the EXAONE 4.0 32B language model.
- Architecture and training: Architectural and training innovations include GQA, 2D RoPE, MTP, multi-stage pre-training, tailored preference optimization, and multimodal RL.
- Capabilities: EXAONE 4.5 demonstrates robust multimodal data processing and strong reasoning and instruction-following abilities.
- Evaluation: Extensive evaluations report highly competitive, state-of-the-art performance across diverse vision and language tasks.
- Implications: As an open-weight model, EXAONE 4.5 is intended to accelerate community-driven research and support future VLA model development.
A Contributors
The report lists authors and contributors, with authors ordered alphabetically by last name.
- All authors are listed in alphabetical order by last name.
- The report identifies a group of core contributors.
- The report separately identifies additional contributors.
B Model License
The agreement defines the EXAONE AI Model broadly and identifies the parties responsible for its use and licensing. It also treats modifications and generated materials as covered concepts.
- Downloading, installing, copying, or using the Model constitutes acceptance of the binding Agreement.
- The Model includes software, algorithms, machine-learning models, related components, and subsequent updates or improvements.
- Derivatives encompass modifications to the Model’s architecture, parameters, data processing methods, functionality, or output.
- Output includes data, results, content, predictions, analyses, insights, and other materials generated directly or indirectly by the Model or Derivatives.
- The Licensor owns, develops, and provides the EXAONE AI Model and grants licenses under the Agreement’s terms.
- The Licensee may be an individual or organization using or intending to use the Model and must ensure authorized users comply with the Agreement.
2. License Grant
The Agreement grants a limited, revocable license for worldwide research and educational use, including evaluation, modification, publication, and distribution under specified conditions.
- The license is limited, non-exclusive, non-transferable, worldwide, and revocable.
- Research and educational use includes evaluation, testing, experimentation, teaching, training, and non-commercial competition participation.
- Licensees may publicly disclose research results and findings, including through papers or presentations.
- Licensees may modify the Model and create Derivatives exclusively for research and educational purposes.
- Modified Models must include “EXAONE” at the beginning of their names.
- Distribution of the Model and Derivatives requires including a copy of the Agreement, while uses outside its scope require written Licensor approval.
3. Restrictions
The Agreement restricts commercial, unlawful, reverse-engineering, and harmful uses of the Model and Derivatives while preserving Licensor ownership and requiring attribution.
- Commercial use of the Model, Derivatives, or Output is prohibited without a separate commercial license.
- Reverse engineering, decompilation, disassembly, and circumvention of technological protections are prohibited except where applicable law permits them.
- The Model and Derivatives may not be used for illegal, fraudulent, unauthorized, or otherwise unlawful activities.
- Ethical Use: Ethical-use rules prohibit false or harmful information, discriminatory or abusive content, rights violations, and conduct causing physical, mental, emotional, or financial harm.
- Intellectual Property: All rights in the Model, modifications, Derivatives, and documentation remain with the Licensor, while the Licensor claims no rights in Output.
- Attribution: Publications and presentations using the Model must provide appropriate attribution, including the Model’s name and version.
5. No Warranty
The Agreement provides the Model, Derivatives, and Output without warranties and places performance risk and evaluation responsibility on the Licensee. It also limits Licensor liability, permits termination, and preserves selected obligations afterward.
- The Model, Derivatives, and Output are provided “as-is” and “as-available” without express, implied, or statutory warranties.
- The Licensor does not guarantee that operation will meet requirements, remain uninterrupted or error-free, or have defects corrected.
- Licensees use the Model, Derivatives, and Output at their own risk and acknowledge possible bugs, errors, and other limitations.
- The Licensor does not endorse or certify results, conclusions, or recommendations, leaving accuracy and suitability evaluation to the Licensee.
- The Agreement excludes Licensor liability for specified special, incidental, indirect, consequential, exemplary, or punitive damages to the fullest extent permitted by law.
- The Licensee must indemnify the Licensor and related parties for claims and losses arising from use, Agreement violations, or legal violations.
- The Licensor may terminate the Agreement and revoke use rights for breach, after which the Licensee must stop using and destroy Model and Derivative copies.
- Ownership, no-warranty, liability-limitation, and termination provisions survive termination.
8. Governing Law
The agreement is governed by Korean law, disputes are resolved through KCAB arbitration, and the Licensor may unilaterally modify the terms. Use of the EXAONE AI Model signifies acceptance, while the agreement supersedes prior arrangements.
- The agreement is governed by and construed under the laws of the Republic of Korea.
- Disputes relating to the agreement are subject to final arbitration administered by KCAB under its International Arbitration Rules.
- The Licensor may modify the agreement at its sole discretion, with changes effective upon posting or other communication.
- The agreement supersedes prior agreements and excludes inconsistent or additional Licensee purchase-order terms.
- Downloading, installing, or using the EXAONE AI Model constitutes acknowledgment and acceptance of the agreement.