/>

Introduction

The global development of large language models (LLMs) is currently experiencing intense competition, with leading countries striving to deploy models with superior performance. In this race, closed-source models currently hold a competitive advantage, while open-weight models are rapidly catching up by employing aggressive scaling strategies. A major factor behind the momentum of open-weight models is the effectiveness of scaling in terms of model size, which has now surpassed hundreds of billions of parameters and is approaching the trillion-parameter scale. The scaling effort is crucial in reducing the performance gap between closed-source and open-weight models.

However, the situation in South Korea presents unique challenges. Compared to global leaders, Korea faces relative shortages in AI-specialized data centers and AI chips, which have limited the development of large-scale models. As a result, previous efforts have focused on cost-effective smaller-scale models (on the order of tens of billions of parameters). Despite these challenges, building a robust and reliable foundation for AI transformation fundamentally requires acquiring a model that demonstrates top-tier performance on a global scale. To address this infrastructure gap, the Korean government has initiated a strategic program aimed at providing essential resources—such as GPUs—for the development of large-scale AI models. LG AI Research has actively participated in this initiative, leveraging government support to develop the K-EXAONE foundation model, which is detailed in this technical report.

K-EXAONE builds on the hybrid architecture of EXAONE 4.0 , combining reasoning and non-reasoning capabilities to enhance both general-purpose and specialized use cases. It also uses a hybrid attention mechanism that integrates global and local attention modules, enabling efficient processing of long-context inputs—a critical feature for real-world applications.

A key architectural innovation that sets K-EXAONE apart is the adoption of the Mixture-of-Experts (MoE) paradigm, a design increasingly used in the state-of-the-art models, which allows for scalable and efficient computation. Additionally, while EXAONE 4.0 supports Korean, English, and Spanish, K-EXAONE extends multilingual coverage by enhancing the tokenizer to include German, Japanese, and Vietnamese, thereby broadening its applicability across diverse linguistic contexts.

Modeling

Model Configurations

K-EXAONE is architecturally distinct from the EXAONE series previously released by LG AI Research. While EXAONE adopts a dense modeling paradigm, K-EXAONE is designed with a MoE architecture, which enables resource-efficient scaling of model capacity and has been increasingly adopted for training models at the 100B-parameter scale and beyond.

As illustrated in Figure 2, K-EXAONE employs a fine-grained sparse MoE design inspired by prior work , consisting of 128 experts, where the top-8 experts are activated per token together with an additional shared expert, resulting in nine concurrently active experts per routing decision. Although the total number of parameters amounts to 236B, only approximately 23B parameters are activated, enabling high representational diversity and strong performance while maintaining resource-efficient training and inference. To improve routing stability and expert utilization efficiency, sequence-level load balancing is employed in the MoE routing mechanism, and a dropless routing policy  is adopted to ensure that all tokens are dispatched to experts without capacity-based dropping, thereby stabilizing gradient flow and improving convergence behavior in large-scale MoE training.

/>

In addition, K-EXAONE integrates a dense-layer-based Multi-Token Prediction (MTP) module  to enable resource-efficient auxiliary training, minimizing routing overhead and memory consumption while enhancing future-token predictive capability. During inference, K-EXAONE leverages the MTP block for self-drafting, achieving an approximately 1.5$`\times`$ improvement in decoding throughput compared to standard autoregressive decoding.

K-EXAONE supports a maximum context length of 256K tokens and incorporates the hybrid attention architecture originally introduced in EXAONE 4.0, which significantly reduces memory consumption and computational overhead compared to full global attention (GA) across all layers, enabling cost-efficient long-context modeling.

To enhance training stability and long-context extrapolation, K-EXAONE incorporates two architectural features—QK Norm and SWA (Sliding Window Attention) -only RoPE (Rotary Positional Embeddings)—inherited from EXAONE 4.0. QK Norm applies layer normalization to the query and key vectors prior to attention computation, mitigating attention logit explosion in deep networks and stabilizing training dynamics, while RoPE are selectively applied only to SWA layers, preventing interference with global token interactions and improving robustness to long-sequence extrapolation.

To further optimize inference efficiency under long-context settings, the sliding-window size is reduced from 4,096 to 128, thereby minimizing KV-cache usage while preserving modeling capacity. In detail, among various architectural designs for efficient long-context inference, K-EXAONE adopts SWA and GA, both of which are natively supported by modern LLM inference engines, improving deployment accessibility and system compatibility. The detailed model configuration is presented in Table 1.

Tokenizer

Compared to previous models in EXAONE series, we redesign the tokenizer and increase the vocabulary size from 100K to 150K to improve token efficiency, downstream task performance, and multilingual scalability. We retain the 70% high-frequency portion of the original vocabulary and reallocate capacity to expand coverage for additional languages, STEM (Science, Technology, Engineering, Mathematics), and code domains. To further improve efficiency, we adopt SuperBPE  strategy that introduces superword tokens, allowing common word sequences to a single token and reducing overall sequence length. The superword tokens make up about 20% of the K-EXAONE’s vocabulary, allocated across English, Korean, and multilingual coverage in a 2:3:1 ratio.

r0.43

In addition, we update the pre-tokenization regex (regular expression) and normalization method to support the expanded vocabulary and superword units. We replace pre-tokenization regex for handling superword boundaries, line breaks, and multilingual Unicode characters. We also switch Unicode normalization  from NFKC to NFC to preserve semantic distinctions in superscripts, subscripts, and symbol-rich text commonly found in code and STEM corpora. Figure [fig:tokenizer_efficiency] shows tokenizer efficiency measured as bytes per token, where higher values indicate that each token represents a larger span of text in bytes. K-EXAONE consistently improves efficiency across English, Korean, multilingual, STEM, and code inputs, achieving an approximately 30% improvement on average over the previous EXAONE tokenizer.

Training

Pre-training

K-EXAONE utilizes a strategic three-stage pre-training curriculum to progressively build foundational knowledge, domain expertise, and reasoning capabilities. While inheriting the data pipeline of EXAONE 4.0, we apply multi-faceted data filtering process to ensure high-quality data. In addition, we extend the model’s linguistic coverage to include German, Japanese, and Vietnamese. Furthermore, we synthesize the corpora with reasoning trajectories to better support post-training. The total amount of data and computational resources used for pre-training are summarized in Table 2.

Extending Multilingual Coverage

To broaden multilingual support, we expand the language coverage of EXAONE 4.0 beyond English, Korean, and Spanish to include German, Japanese, and Vietnamese. We incorporate high-quality web text in these additional languages. Since pre-training data distribution varies substantially across languages, we mitigate this imbalance through targeted data synthesis. Leveraging cross-lingual knowledge transfer, we generate synthetic corpora that propagate specialized knowledge and reasoning patterns across languages, ensuring a balanced knowledge distribution and consistent performance regardless of the input language.

Thinking-Augmented Data Synthesis

We further enrich our data synthesis pipeline by extending the curation strategy of EXAONE 4.0 to incorporate explicit reasoning supervision. Motivated by , we generate document-grounded thinking trajectories and combine them with the source content into unified samples that encode step-by-step inference. These thinking-augmented corpora facilitate the transfer of reasoning behaviors and improve the effectiveness of subsequent post-training.

Training Setup

K-EXAONE is natively trained with FP8 precision and achieves training loss curves comparable to those obtained under BF16 precision, demonstrating that FP8 training preserves optimization stability while enabling full quantization-aware convergence. We adopt the Muon optimizer for all training stages in conjunction with the Warmup–Stable–Decay (WSD) learning rate scheduler. The maximum learning rate is set to $`3.0 \times 10^{-4}`$, with a linear warm-up phase followed by a stable plateau and a subsequent decay schedule. For MoE regularization, the sequence auxiliary loss coefficient is fixed to $`1.0 \times 10^{-4}`$, and the expert bias update factor is also set to $`1.0 \times 10^{-4}`$ throughout training. For the MTP objective, a loss weight of 0.05 is applied.

Context Length Extension

K-EXAONE is designed to support a maximum context length of 256K tokens. To achieve this capability, we employ a two-stage context length extension procedure. The base model is pre-trained with a maximum context length of 8K tokens, and is subsequently extended (i) from 8K to 32K tokens in the first stage, and (ii) from 32K to 256K tokens in the second stage. Across both stages, we preserve the same high-level data-mixture components, Rehearsal Dataset, Synthetic Reasoning Dataset, and End-to-end Long-Document Dataset, while adjusting their sampling ratios to match the target context regime and the stability requirements of each stage.

Rehearsal Dataset for Preserving Short-Context Performance

A primary risk of long-context specialization is degradation of short-context performance. To mitigate this, we incorporate a Rehearsal Dataset as a core component during context extension. The Rehearsal Dataset reuses high-quality samples drawn from the pre-training distribution and other short-context data, providing a consistent training signal that anchors the model’s behavior in shorter regimes. We include rehearsal in both Stage 1 (8K$`\rightarrow`$32K) and Stage 2 (32K$`\rightarrow`$256K), while adjusting its proportion stage-wise to ensure that long-context learning signals are adequately incorporated. This design helps preserve the short-context baseline after context extension, as verified by standard short-context benchmarks and internal validation metrics.

Synthetic Reasoning Dataset for Boosting Reasoning Capability

To strengthen reasoning performance, we additionally train on a Synthetic Reasoning Dataset. This dataset comprises challenging problems in mathematics, science, and competitive programming, and includes synthesized reasoning content that encourages the model to learn intermediate reasoning patterns beyond final answers. The overall objective is aligned with prior synthetic reasoning approaches: improving the model’s consistency and robustness in multi-step reasoning. We integrate this dataset throughout context extension so that the model improves reasoning quality even under long inputs.

Long-Document Dataset for Long-Range Adaptation and Verification with NIAH

To ensure strong long-context performance, we place particular emphasis on a Long-Document Dataset during the extension phases, consisting of full-document sequences that can be consumed within a single training instance. We train on these samples in an end-to-end manner by feeding entire long-document sequences without truncation, encouraging the model to capture long-range dependencies. Stage 1 prioritizes stable performance up to 32K tokens, whereas Stage 2 increases the emphasis on long-document samples to better model dependencies up to 256K tokens. To systematically monitor potential performance degradation, we conduct (i) short-context evaluations following the same protocol used in pre-training, and (ii) Needle-In-A-Haystack (NIAH)  tests to assess the model’s ability to retain and retrieve information from long contexts. Training is iteratively repeated until the model consistently achieves near-perfect NIAH performance across the target context ranges for each stage (“green light”), indicating that K-EXAONE successfully extends to 256K tokens without compromising overall performance.

Post-training

In K-EXAONE, the post-training process is primarily organized into three stages: (i) large-scale supervised fine-tuning (SFT) to learn to follow a variety of user instructions and produce corresponding responses, (ii) reinforcement learning (RL) on reasoning-intensive and verifiable tasks, and (iii) preference learning to align with human preferences. Most of the dataset generation pipelines follow those of EXAONE 4.0. To summarize, we first categorize instruction-following tasks into several domains and adopt distinct generation methods or experts. To enhance Korean-specific capabilities, we utilize public and institutional data provided by Ministry of Science and ICT (MSIT) of South Korea and its affiliated agencies, such as National Information Society Agency (NIA)¹ and Korea Data Industry Promotion Agency (K-DATA)².

Training Agentic Tool Use

Aggregating or constructing real-world agentic tool environments is costly and inefficient. Instead, we leverage LLMs to build synthetic tool environments, including tool-use scenarios and verifiable pass criteria for various tasks (e.g., coding or general tool-calling scenarios). We then evaluate the LLMs on the generated environments to filter out unrealistic and unsolvable cases. Through this process, we obtain hundreds of verifiable, realistic tool-use tasks, along with their corresponding evaluation environments.

Enabling Web Search with Sub-Agents

When K-EXAONE performs web search as the primary agent, we augment it with two sub-agents: a summarizer and a trajectory compressor. The summarizer sub-agent distills fetched webpages so that K-EXAONE can avoid processing long and noisy web text. Once the tool-calling history exceeds a predefined number of steps, the trajectory compressor compresses the full interaction into a single JSON-formatted structured record that captures key facts from tool outputs and the remaining open questions to investigate. This design improves context efficiency by preventing redundant tool results from being repeatedly exposed to K-EXAONE. At inference time, both sub-agents are implemented using the same underlying model as K-EXAONE.

Reinforcement Learning

To enhance the reasoning capability of the post-trained model, we perform reinforcement learning (RL) with verifiable rewards. We train in a multi-task setup covering math, code, STEM, and instruction-following. For verification, we use a combination of rule-based verifiers and an LLM-as-a-judge. For optimization, we use AGAPO  with an off-policy policy-gradient using truncated importance sampling.

For training efficiency at scale, we adopt an off-policy policy-gradient objective with truncated importance sampling, following prior work . We also apply zero-variance filtering by dropping prompts whose sampled rollouts receive identical rewards, resulting in zero advantages. We employ both group-level advantage computation and global advantage normalization to capture both within-group relative reward signals and batch-level information. During training, we exclude the KL penalty to improve performance while avoiding unnecessary computation. Finally, we freeze the MoE router throughout RL training.

The RL objective is defined for a question $`q\sim P(Q)`$. For each question, we sample a group of $`G`$ candidate responses $`O=\{o_1,\dots,o_G\}`$ from a rollout policy $`\pi_{\theta_{\mathrm{rollout}}}`$, and assign each response a verifiable reward $`r_i\in[0,1]`$. We write each response as a token sequence $`o_i=(o_{i,1},\dots,o_{i,|o_i|})`$. We apply truncated importance sampling at the token level with a stop-gradient function $`\mathbf{sg}(\cdot)`$.

\begin{equation}
\mathcal{J}_{\mathrm{AGAPO}}(\theta)=
\operatorname*{\mathbb{E}}_{\substack{
q\sim P(Q),\\
\{o_i\}_{i=1}^{G}\sim\pi_{\theta_{\mathrm{rollout}}}(O\mid q)
}}
\Biggl[
  \frac{1}{G}\sum_{i=1}^{G}
    \Bigl(
      \frac{1}{|o_i|}\sum_{t=1}^{|o_i|}
      \mathbf{sg}\!\bigl(\min(\rho_{i,t},\,\epsilon)\bigr)\,
      A_{\mathrm{global},i}\,
      \log\pi_{\theta}(o_{i,t}\mid q, o_{i,<t})
    \Bigr)
\Biggr].
\end{equation}

\begin{equation}
\rho_{i,t}
=
\frac{\pi_{\theta}(o_{i,t}\mid q, o_{i,<t})}{\pi_{\theta_{\mathrm{rollout}}}(o_{i,t}\mid q, o_{i,<t})}.
\quad
A_{\mathrm{group},i}=r_i-\frac{1}{G-1}\sum_{j\neq i} r_j,
\quad
A_{\mathrm{global},i}=
\frac{A_{\mathrm{group},i}-\operatorname{mean}\!\bigl(\{A_{\mathrm{group},k}\}_{k}\bigr)}
{\operatorname{std}\!\bigl(\{A_{\mathrm{group},k}\}_{k}\bigr)}.
\end{equation}

Preference Learning

After RL training, we perform a preference learning stage to better align the model with human preferences. In this stage, we aim to preserve reasoning performance while focusing training on general alignment domains such as chat, safety, instruction-following, agentic tool use, and creative writing. We propose GrouPER (Group-wise SimPER), an improved variant of SimPER , and show that it improves general-domain performance.

Inspired by the GRPO , GrouPER samples multiple responses for each query and trains the model using group-wise advantages. For each response, we compute a preference reward by combining rule-based rewards with rubric-based generative rewards that score responses along multiple dimensions. We then compute a group-level advantage from these scores and integrate it into the SimPER-style objective.

For each input $`x\sim P(X)`$, we sample a group of $`G`$ candidate responses $`O=\{o_1,\dots,o_G\}`$ from the initial policy $`\pi_{\theta_{\mathrm{init}}}`$. Each response $`o_i`$ is assigned a preference reward $`r_{\mathrm{pref},i}\in\mathbb{R}`$. We compute a group-level advantage by (i) standardizing $`r_{\mathrm{pref}}`$ and (ii) scaling it to $`[-1,1]`$. The objective function minimizes the following:

\begin{equation}
\mathcal{L}_{\mathrm{GrouPER}}(\theta)=
-\operatorname*{\mathbb{E}}_{\substack{
x\sim P(X),\\
\{o_i\}_{i=1}^{G}\sim\pi_{\theta_{\mathrm{init}}}(O\mid x)
}}
\Biggl[
  \frac{1}{G}\sum_{i=1}^{G}
  \Bigl(
    A_{\mathrm{pref},i}\;
    \exp\!\Bigl(\frac{1}{|o_i|}\log \pi_{\theta}(o_i\mid x)\Bigr)
  \Bigr)
\Biggr].
\end{equation}

\begin{equation}
z_i=
\frac{
r_{\mathrm{pref},i}-\operatorname{mean}\!\bigl(\{r_{\mathrm{pref},j}\}_{j=1}^{G}\bigr)
}{
\operatorname{std}\!\bigl(\{r_{\mathrm{pref},j}\}_{j=1}^{G}\bigr)
},
\quad
A_{\mathrm{pref},i}
=
2\cdot
\frac{
z_i-\min\!\bigl(\{z_j\}_{j=1}^{G}\bigr)
}{
\max\!\bigl(\{z_j\}_{j=1}^{G}\bigr)-\min\!\bigl(\{z_j\}_{j=1}^{G}\bigr)
}
-1
\in[-1,1].
\end{equation}

Data Compliance

Developing AI models requires a large amount of data, and the acquisition and utilization of this data can lead to various legal issues, such as copyright infringement, intellectual property infringement, and personal information protection violations. To minimize these risks, LG AI Research conducts AI Compliance reviews throughout the entire process of data collection, AI model training, and information provision. For more detailed information, please refer to the EXAONE 3.0 Technical Report  and the LG AI Ethics Principles .

Evaluation

Benchmarks and Setup

We evaluate K-EXAONE on a diverse set of benchmarks spanning nine categories below:

• World Knowledge: MMLU-Pro , GPQA-Diamond , and Humanity’s Last Exam³ 

• Math: IMO-AnswerBench , AIME 2025 , and HMMT Nov 2025 

• Coding / Agentic Coding: LiveCodeBench Pro (25Q2 Medium) , LiveCodeBench v6 ,
  Terminal-Bench 2.0 , and SWE-bench Verified 

• Agentic Tool Use: $`\tau^2`$-Bench  and BrowseComp 

• Instruction Following: IFBench  and IFEval 

• Long Context Understanding: AA-LCR  and OpenAI-MRCR 

• Korean: KMMLU-Pro , KoBALT , CLIcK , HRM8K , and Ko-LongBench (in-house)

• Multilinguality⁴: MMMLU  and WMT24++ 

• Safety: WildJailbreak  and KGC-Safety⁵ (in-house)

To evaluate our model, we set the temperature to 1.0 and top-p  to 0.95. We set the context length to 160K for the Long Context Understanding benchmarks, while 128K for others. We disable the MTP at inference time. For baseline models, when official scores are unavailable, we evaluate them in our internal environment with inference parameters set to the recommended configuration for each model. Please refer to Appendix 10 for the detailed evaluation setup of each benchmark and Appendix 11 for the in-house benchmarks.

Results

Table [tab:results_reasoning] and [tab:results_non_reasoning] present the benchmark results of K-EXAONE in Reasoning and Non-Reasoning modes, respectively.

Reasoning Abilities

As shown in Table [tab:results_reasoning], K-EXAONE achieves competitive performance and often leads across the evaluated tasks. In world knowledge benchmarks, such as MMLU-Pro, GPQA-Diamond, Humanity’s Last Exam, K-EXAONE demonstrates competitive academic knowledge understanding and reasoning capabilities. Additionally, it surpasses gpt-oss-120b and Qwen3-235B-A22B-Thinking-2507 in all mathematics benchmarks, except for HMMT Nov 2025 against the Qwen model. For coding, K-EXAONE outperforms most compared baselines on the competitive programming benchmark LiveCodeBench v6 and shows comparable performance on LiveCodeBench Pro. Beyond competitive programming, K-EXAONE shows improved performance over its predecessor on the in-house CodeUtilityBench, effectively combining algorithmic reasoning with practical coding abilities for real-world coding workflows. See Appendix 11.1 for further details.

Agentic Abilities

We further assess K-EXAONE’s agentic abilities in settings that require goal-directed, multi-step interaction and tool use. On agentic coding benchmarks Terminal-Bench 2.0 and SWE-bench Verified, K-EXAONE attains 29.0 and 49.4, indicating its potential in realistic software development workflows. On agentic tool use benchmarks $`\tau^2`$-Bench, it achieves 73.2 (the weighted average score), suggesting reliable tool selection and effective information seeking over multi-step interactions.

General Abilities

To evaluate the model’s general abilities, we use widely adopted open-source benchmarks covering instruction following and long-context understanding. For instruction following, K-EXAONE achieves scores of 67.3 on IFBench and 89.7 on IFEval in the reasoning mode, outperforming the majority of competing models. Additionally, regarding long-context understanding, K-EXAONE scores 53.5 on AA-LCR and 52.3 on OpenAI-MRCR in the reasoning mode, demonstrating competitive performance and robust scaling with longer contexts. In the non-reasoning mode, it achieves 45.2 on AA-LCR and 65.9 on OpenAI-MRCR, surpassing strong baselines by a large margin and indicating efficient, accurate long-context processing.

Korean and Multilingual Abilities

Across Korean-centric benchmarks, K-EXAONE shows strong performance among open-weight reasoning models: 67.3 on KMMLU-Pro (professional knowledge), 61.8 on KoBALT (advanced linguistic competence), 83.9 on CLIcK (linguistic and cultural competence), 90.9 on HRM8K (olympiad-level math reasoning), and 86.8 on Ko-LongBench (long-context understanding; see Appendix 11.2). Overall, these results indicate competitive Korean professional knowledge, language competence, mathematical reasoning, and long-context capability.

For multilingual evaluation, we report performance averaged over the non-English supported languages. K-EXAONE demonstrates competitive multilingual knowledge understanding scoring 85.7 on MMMLU. Translation performance is assessed using WMT24++, where the model attains an average score of 90.5, indicating stable multilingual translation quality. We describe performance per language in Appendix 12.1

Safety

The model achieves competitive performance on both WildJailbreak, designed to assess robustness to a wide range of harmful prompts, and KGC-Safety, which is designed to jointly evaluate Korean sociocultural contexts and global ethical standards. This suggests that K-EXAONE effectively mitigates risks and handles sensitive queries without incurring significant performance trade-offs on general downstream tasks.

Limitations

K-EXAONE language models, like all existing language models, have certain limitations and may occasionally generate inappropriate responses. The language model generates responses based on the output probability of tokens, and it is determined during learning from training data. While we make every effort to exclude personal, harmful, and biased information from the training data, some problematic content may still be included, potentially leading to undesirable responses. Please note that the text generated by K-EXAONE language models does not reflect the views of LG AI Research.

• Inappropriate answers may be generated, which contain personal, harmful or other inappropriate information.

• Biased responses may be generated, which are associated with age, gender, race, and so on.

• The generated responses rely heavily on statistics from the training data, which can result in the generation of semantically or syntactically incorrect sentences.

• Since the models do not reflect the latest information, the responses may be false or contradictory.

LG AI Research strives to reduce potential risks that may arise from K-EXAONE language models. Users are not allowed to engage in any malicious activities (e.g., keying in illegal information) that may induce the creation of inappropriate outputs violating LG AI’s ethical principles when using K-EXAONE language models.

Deployment

Section 9 in the Appendix provides license information for using the K-EXAONE models. Understanding the license information is essential for the legal utilization of the language model.

Conclusion

The development of K-EXAONE represents a significant advancement in AI technology. By adopting a MoE architecture, K-EXAONE achieves efficient scaling while maintaining high performance. The integration of a hybrid attention mechanism enables the model to effectively handle long-context inputs and outputs, a critical feature for complex tasks. K-EXAONE extends its multilingual support to include Korean, English, Spanish, German, Japanese, and Vietnamese, making it highly versatile across diverse linguistic contexts.

The training process of K-EXAONE is rigorous, involving a comprehensive data curation and synthesis pipeline, a three-stage curriculum, and FP8 precision training. These methods can inject parametric knowledge and ensure stable convergence. Notably, K-EXAONE supports a maximum context length of 256K tokens, achieved through a two-stage context length extension procedure. Post-training, K-EXAONE undergoes SFT, RL with verifiable rewards, and preference learning to align the model closely with human preferences. This alignment process ensures that the model behaves in a manner that is both ethical and user-friendly.

Performance evaluations across various benchmarks demonstrate that K-EXAONE excels in reasoning, agentic capabilities, general knowledge, multilingual understanding, and long-context processing. These results underscore its competitiveness in the AI field, showing that progress in AI benefits both individuals and society, thereby contributing to advancing AI for a better life.

Contributors

All authors are listed in alphabetical order by last name.

Core Contributors

Eunbi Choi, Kibong Choi, Seokhee Hong, Junwon Hwang, Hyojin Jeon, Hyunjik Jo, Joonkee Kim, Seonghwan Kim, Soyeon Kim, Sunkyoung Kim, Yireun Kim, Yongil Kim, Haeju Lee, Jinsik Lee, Kyungmin Lee, Sangha Park, Heuiyeen Yeen

Contributors

Hwan Chang, Stanley Jungkyu Choi, Yejin Choi, Jiwon Ham, Kijeong Jeon, Geunyeong Jeong, Gerrard Jeongwon Jo, Yonghwan Jo, Jiyeon Jung, Naeun Kang, Dohoon Kim, Euisoon Kim, Hayeon Kim, Hyosang Kim, Hyunseo Kim, Jieun Kim, Minu Kim, Myoungshin Kim, Unsol Kim, Youchul Kim, YoungJin Kim, Chaeeun Lee, Chaeyoon Lee, Changhun Lee, Dahm Lee, Edward Hwayoung Lee, Honglak Lee, Jinsang Lee, Jiyoung Lee, Sangeun Lee, Seungwon Lim, Solji Lim, Woohyung Lim, Chanwoo Moon, Jaewoo Park, Jinho Park, Yongmin Park, Hyerin Seo, Wooseok Seo, Yongwoo Song, Sejong Yang, Sihoon Yang, Chang En Yea, Sihyuk Yi, Chansik Yoon, Dongkeun Yoon, Sangyeon Yoon, Hyeongu Yun

Model License

K-EXAONE AI Model License Agreement

This License Agreement (“Agreement”) is entered into between LG Management Development Institute Co., Ltd. (“Licensor”) and you (“User”) and governs the use of the K-EXAONE AI Model (“Model”). By downloading, installing, copying, or using the Model, you agree to comply with and be bound by the terms of this Agreement. If you do not agree to all terms, you must not download, install, copy, or use the Model. This Agreement constitutes a binding legal contract between User and Licensor.

1. Definitions

1.1 “Model” means the artificial intelligence model provided by Licensor, including all software, algorithms, machine learning models, or related components provided by Licensor, together with all updates, improvements, enhancements, bug fixes, patches, or other modifications thereto implemented automatically or manually.

1.2 “Derivative Work” means any modification, alteration, improvement, enhancement, adaptation, or derivative work of the Model created by User or a third party, including any changes to the Model’s architecture, parameters, data processing methods, or any other aspect of the Model that modifies its functionality or output.

1.3 “Output” means all data, results, content, predictions, analyses, insights, or other materials generated by the Model or Derivatives Work, whether in their original form or further processed or modified by User. This includes, but is not limited to, text or numerical data generated directly or indirectly through the use of the Model.

1.4 “Licensor” means the provider that lawfully offers the K-EXAONE AI Model. Licensor retains all rights to the Model and has the right to grant a license for its use under the terms specified in this Agreement.

1.5 “User” means an individual, organization, corporation, academic institution, government agency, or other entity that uses or intends to use the Model under the terms of this Agreement. User is responsible for ensuring that all authorized users accessing or using the Model on its behalf comply with this Agreement.

2. License Grant

2.1 License Grant: Subject to the terms and conditions set forth in this Agreement and Section 2.2, Licensor grants to the User a non-exclusive, non-transferable, worldwide, irrevocable license to access, download, install, modify, use, distribute, and create derivative works of the Model for commercial and non-commercial purposes. In the event the Model or Derivative Work is distributed, this Agreement shall be distributed alongside it to ensure the license terms are maintained, and the name of the Model and Derivative Work shall begin with “K-EXAONE”.

2.2 Distribution, sublicensing, or making the Model and Derivative Work available to third parties for commercial purposes requires separate agreement with Licensor.

3. Exceptions and Restrictions

3.1 Reverse Engineering: Except as expressly permitted by applicable law, User shall not attempt to decompile, disassemble, reverse engineer, or derive the source code, underlying ideas, algorithms, or structure of the Model. Any attempts to circumvent or evade any technical protection measures applied to the Model are strictly prohibited.

3.2 Illegal Use: User shall not use the Model or Derivative Work for any illegal, fraudulent, or unauthorized activities, or for purposes that violate applicable laws or regulations, including but not limited to, the creation, distribution, or dissemination of malicious, deceptive, or illegal content.

Ethical Use: User shall ensure that the Model or Derivative Work is used ethically and responsibly in compliance with the following guidelines:

a. Model and Derivative Work must not be used to generate, disseminate, or amplify false, misleading, or harmful information, including fake news, misinformation, or inflammatory content.

b. Model and Derivative Work must not be used to create, distribute, or promote content that is discriminatory, harassing, defamatory, insulting, or otherwise offensive toward individuals or groups based on race, gender, sexual orientation, religion, nationality, or other protected characteristics.

c. Model and Derivative Work must not infringe upon the rights of others, including intellectual property rights, privacy rights, or other rights recognized by law. User must obtain all necessary permissions and consents before using the Model and Derivative Work in a manner that could affect the rights of third parties.

d. Model and Derivative Work must not be used in a manner that causes physical, mental, emotional, or financial harm to any individual, organization, or community. User must take all reasonable measures to prevent the misuse or abuse of the Model and Derivative Work that could result in harm or injury.

4. Ownership

4.1 Intellectual Property Rights: User acknowledges that this Agreement does not transfer to the User any ownership or patent rights related to the Model or any trademarks, service marks, and logos.

4.2 Output: Licensor claims no ownership over any output generated by the Model or Derivative Work, and the use of such output is solely the responsibility of User.

5. Warranty

5.1 Provided “As Is”: The Model and Derivative Work are provided “as is,” without any warranty or representation of any kind, whether express, implied, or statutory. Licensor disclaims all warranties, including but not limited to implied warranties of merchantability, fitness for a particular purpose, accuracy, reliability, and non-infringement, as well as any warranties arising from course of dealing or trade usage.

5.2 Performance and Reliability: Licensor does not warrant or guarantee that the Model or Derivative Work will meet User’s requirements, that the operation of the Model or Derivative Work will be uninterrupted or error-free, or that defects in the Model will be corrected. User acknowledges that use of the Model or Derivative Work is at their own risk and that the Model or Derivative Work may contain bugs, errors, or other limitations.

5.3 Warranty Disclaimer: Licensor does not warrant, endorse, or certify any results, conclusions, or recommendations arising from the use of the Model. User bears sole responsibility for evaluating the Model’s accuracy, reliability, and suitability for its intended purpose.

6. Limitation of Liability

6.1 Indemnity for Damages: To the maximum extent permitted by applicable law, Licensor shall not be liable for any special, incidental, indirect, consequential, punitive, or exemplary damages, including the loss of business profits, business interruption, loss of business information, data loss, or any other pecuniary or non-pecuniary loss arising from the use or inability to use the Model, Derivative Work, or Outputs, even if Licensor has been advised of the possibility of such damages.

6.2 Indemnification: User agrees to indemnify, defend, and hold harmless Licensor, its affiliates, officers, directors, employees, and agents from and against any and all claims, liabilities, damages, losses, costs, or expenses (including reasonable attorneys’ fees) arising out of or in connection with your use of the Model, Derivative Work, or Outputs, including any breach of this Agreement or applicable law.

7. Termination

7.1 Termination by Licensor: Licensor reserves the right to terminate this Agreement and revoke the User’s right to use the Model at any time, with or without cause and without prior notice, if User breaches any term of this Agreement. Termination shall be effective immediately upon notice.

7.2 Effect of Termination: Upon termination of this Agreement, User shall immediately cease all use of the Model and Derivative Work and destroy all copies of the Model and Derivative Work in the User’s possession or control, including any backup or archival copies. User shall provide written proof to Licensor that such destruction has been completed.

7.3 Survival: The provisions of this Agreement that by their nature should survive termination (including, without limitation, Section 4 (Ownership), Section 5 (Warranty), Section 6 (Limitation of Liability), and this Section 7 (Termination)) shall survive termination.

8. Governing Law

8.1 Governing Law: This Agreement shall be construed and governed by the laws of the Republic of Korea, without giving effect to its conflict of laws principles.

8.2 Dispute Resolution: All disputes, controversies, or claims arising out of or in connection with this Agreement, including its existence, validity, interpretation, performance, breach, or termination, shall be finally settled by arbitration administered by the Korea Commercial Arbitration Board (KCAB) in accordance with the KCAB International Arbitration Rules in effect at the time of the commencement of the arbitration. The place of arbitration shall be Seoul, Republic of Korea. The arbitral tribunal shall consist of one (1) arbitrator. The language of the arbitration shall be Korean.

9. Miscellaneous

9.1 Entire Agreement: This Agreement constitutes the entire agreement between User and Licensor regarding the subject matter hereof and supersedes all prior oral or written agreements, representations, or understandings. Any terms in a purchase order or other document submitted by the User concerning the Model that add to, differ from, or are inconsistent with the terms of this Agreement shall not be binding upon Licensor and shall be null and void.

By downloading, installing, or using the K-EXAONE AI Model, User acknowledges that they have read and understood the terms of this Agreement and agree to be bound by them.

Evaluation Setup Details

When evaluating models, we try to follow the official evaluation setup for each benchmark. Following is the specific setting we use in our internal evaluation environment. Not mentioned benchmarks are evaluated under official setup.

Multiple-Choice Questions

For multiple-choice questions, we prompt models as in Figure 3 and Figure 4 and parse the final option letter.

Evaluation Prompts

For benchmarks in the math category, we evaluate models with the prompt in Figure 5 and Figure 6, parse the final answer, and compare it with the ground-truth answer through either exact matching or LLM-based equality checking. For IMO-AnswerBench, we use the official evaluation and judging prompts.

Humanity’s Last Exam

In our internal evaluation environment, we use gpt-5-mini-2025-08-07 model as a judge LLM. The judge prompt is from the official.

Terminal-Bench 2.0

When evaluating our model and baselines for which the official scores are unavailable, we use Terminus 2 as the default agent.

SWE-Bench Verified

We evaluate our model using mini-SWE-agent  as our default agent, and use the same setup to obtain scores when official results are unavailable.

BrowseComp

To evaluate K-EXAONE on the BrowseComp benchmark, which is among the most challenging search benchmarks, we adopt the summarizer and the trajectory compressor as described in Section 3.3. We set the maximum number of tool-call steps to 500 and invoke the trajectory compressor every 50 steps. We do not reproduce baseline scores because the setups and pipelines of search agents vary across models. Therefore, we present their scores only if they are reported in the models’ official technical reports or blogs.

IFBench and IFEval

The metrics of IFbench and IFEval benchmarks are prompt-loose and prompt-strict, respectively.

OpenAI-MRCR

We follow the official OpenAI-MRCR protocol, requiring the model to prepend the provided alphanumeric hash. Scores are computed using the difflib.SequenceMatcher ratio. For each context-length bin, we average the scores from the 2-needle, 4-needle, and 8-needle settings to obtain a bin-level score. We evaluate bins up to 128K tokens (despite MRCR supporting contexts up to 1M) and report the macro-average over the resulting bin-level scores.

WMT24++

Figure 7 presents the judging prompt for WMT24++. We adopt the judge prompt from official implementation , but we use gpt-5-mini-2025-08-07 as a judge model. The final scores are average translation scores from LLM judge between en $`\leftrightarrow`$ five non-English supported languages.

WildJailbreak

WildJailbreak results by jointly analyzing ⟨input–model output⟩ pairs using the Qwen3Guard-Gen-8B model  to determine whether responses are safe. Performance is reported using the Safe Rate, defined as the proportion of test cases classified as safe across the full test set, where a higher Safe Rate indicates a safer model.

In-house Benchmarks

Code Utility Benchmark (CodeUtilityBench)

Existing prominent coding benchmarks (e.g., LCB, LCB-Pro) primarily focus on competitive programming problems. However, these benchmarks have limitations in adequately capturing the diverse real-world scenarios in which users employ LLMs for coding tasks. To address this gap, we construct CodeUtilityBench, designed to evaluate the practical performance of LLMs in real-world coding workflows.

CodeUtilityBench is structured around real-world usage patterns and comprises four major categories: (1) Understanding, (2) Implementation, (3) Refinement, and (4) Maintenance. Each category includes four tasks, for a total of 16 tasks:

• Understanding: explain, localize, plan, trace

• Implementation: generate, translate, update, visualize

• Refinement: debug, diff, optimize, verify

• Maintenance: annotate, lint/format, refactor, test

CodeUtilityBench consists of 300 test instances, with 20 instances allocated to each of the 15 tasks (excluding the visualize task). Each test instance includes five evaluation rubrics, which are generated following task-specific protocols. To ensure dataset quality, human experts reviewed all queries and rubrics and replaced unsuitable items. For the visualize task, we use AutoCodeArena , excluding instances from the Problem Solving category as they are not applicable to our task definition.

Evaluation is conducted using an LLM-as-a-judge framework with gpt-5-2025-08-07 as the evaluator. Given $`\langle \textit{query, model response, evaluation rubrics} \rangle`$ as input, the evaluation model outputs five binary labels $`y_1,\dots,y_5 \in \{0,1\}`$, each indicating whether the response satisfies the corresponding rubric. Model performance is reported as a percentage by averaging the per-instance fraction of satisfied rubrics over the entire benchmark.

Figure 8 presents the detailed evaluation results. K-EXAONE achieves an overall score of 71.9%, showing improved performance over its predecessor EXAONE-4.0-32B (63.2%). This suggests improved capability in handling real-world coding workflows. K-EXAONE improves markedly in the Understanding and Implementation categories, reflecting a solid ability to comprehend code context and implement solutions across diverse scenarios. However, it exhibits substantial room for improvement in the Maintenance category, which requires sustained code life-cycle management. This highlights an opportunity for further progress on long-term code maintainability.

style="width:90.0%" />

Ko-LongBench

Ko-LongBench is an in-house benchmark designed to evaluate long-context understanding in Korean. It comprises a diverse set of tasks including Document QA, Story Understanding, Dialogue History Understanding, In-Context Learning, Structured QA, and RAG to assess LLMs’ long-context capabilities in practical settings. Detailed benchmark descriptions, dataset statistics, and representative prompt examples are available in the EXAONE 4.0 Technical Report .

KGC-Safety

Existing frameworks often fail to capture the cultural nuances and context-specific sensitivities of Korean society, leading to limitations in reliability and safety. To address this gap, we propose the Korea-Augmented Universal Taxonomy (K-AUT), an ethical framework that integrates universal ethical principles with Korean sociocultural contexts. Building on this taxonomy, we introduce the Korean Global Civic Safety Benchmark (KGC-Safety). Please refer to Appendix 13.2 for further details of the benchmark.

Further Analysis

Multilingual

K-EXAONE achieves higher performance in average comparable to EXAONE 4.0 in Table [tab:results_reasoning],  [tab:results_non_reasoning]. As shown in  [tab:multi_langwise_scores_mmmlu],  [tab:multi_langwise_scores_wmt24pp], performance gains are evenly distributed across languages, resulting in balanced multilingual capability without pronounced degradation or dominance in any single language.

Safety

Developing a Sovereign AI model for Korea necessitated a fundamental reevaluation of safety policies to address the limitations of existing, predominantly Western-centric AI risk taxonomies. These incumbent frameworks often lack the nuance required to handle the unique cultural sensitivities and specific context of Korean society, resulting in gaps in reliability and safety. To overcome these challenges, we introduce the Korea-Augmented Universal Taxonomy (K-AUT), an ethical framework that incorporates global ethical principles while accounting for the cultural context of Korean society. We also present the Korean Global Civic Safety Benchmark (KGC-Safety), which enables systematic evaluation of how well existing publicly available models adhere to these ethical standards.

Korea-Augmented Universal Taxonomy (K-AUT)

This framework is designed to augment universal human values with regional specificities, thereby ensuring both global acceptability and local reliability. K-AUT categorizes potential harms into 4 primary domains and 226 detailed risk areas. To ensure rigorous enforcement, we implement a strict evaluation protocol where violating even one of the five specific judgment criteria established for each risk area automatically classifies a response as inappropriate.

The architecture of K-AUT, as outlined in the Table 3, is grounded in authoritative sources to balance objectivity with cultural nuance. The domains of “Universal Human Values” and “Social Safety” rely on international consensuses—such as UN declarations—to address threats to life and community cohesion. Distinctively, the “Korean Sensitivity” domain serves as the critical augmentation layer, managing sensitive local issues including historical and geopolitical conflicts. By adhering to domestic laws and verified historical records, this domain minimizes context-unaware hallucinations and ensures the model’s output is legally compliant and culturally accurate. Finally, the “Future Risk” domain incorporates predictive ethics to anticipate challenges from emerging technologies. This approach differentiates our model by systematically integrating regional characteristics with universal ethics, offering a modular and scalable blueprint for building safe and reliable Sovereign AI models globally.

KGC-Safety : Korean Global Civic Safety Benchmark

KGC-Safety consists of 2,260 test instances, constructed by sampling 10 test cases for each of the 226 categories defined in K-AUT. The benchmark supports comprehensive evaluation across multiple problem types, including multilingual scenarios (Korean, English, Spanish, German, Japanese, and Vietnamese), multi-turn, adversarial, and naive settings. Detailed statistics are provided in Table  [tab:kgc_statistics].

Evaluation is conducted using an LLM-as-a-judge framework with gpt-4.1-mini-2025-04-14, which takes ⟨query – model response – evaluation criteria⟩ as input and outputs a binary judgment [0, 1] indicating whether the response satisfies the defined safety standards. Model performance is reported using the Safe Rate, defined as the proportion of test cases classified as safe across the entire benchmark. Detailed evaluation results are provided in Table  [tab:kgc_results].

Overall, most models achieve relatively high Safe Rates in categories such as Universal Human Values and Social Safety, which primarily assess harmful, violent, and bias-related behaviors. In contrast, models tend to exhibit lower Safe Rates in the Future Risk category—designed to evaluate emerging risks associated with technological advances—as well as in Korean Sensitivity, which reflects Korean-specific cultural and social considerations. For a detailed example of the KGC-Safety, refer to Figure  9.