K-EXAONE: Korea’s Flagship Frontier AI Model

Discover K-EXAONE, LG’s 236B-parameter national AI model, built with MoE, 260K-token context, and strong global benchmark scores for math, code, and safety.

K-EXAONE is positioned as South Korea’s flagship national AI foundation model, built by LG AI Research with a Mixture-of-Experts architecture that delivers frontier-level performance while keeping training and inference costs efficient. It combines advanced efficiency techniques, multi-stage training, and strong benchmark results to signal Korea’s push for AI sovereignty alongside leading US and Chinese models.​

Korea’s Push for a National AI Model

As AI races ahead globally, the United States and China are competing to build massive large language models (LLMs) with hundreds of billions to trillions of parameters, raising the bar for performance and strategic capability. Korea, despite strong ambitions, has so far lacked large-scale AI infrastructure such as dedicated data centers and high-performance AI chips, which has limited domestic efforts mostly to models in the billions-of-parameters range. With AI now functioning as economic and security infrastructure rather than just a technology, Korea has reached the stage where it requires a national, world-class foundation model.​

To close this gap, the Korean government launched the “Proprietary AI Foundation Model Project,” providing high-performance GPUs, curated data, and talent to strengthen the country’s AI industry. LG AI Research joined this national initiative and developed K-EXAONE, a large-scale LLM built with proprietary Korean technology that serves as a flagship K-AI model. The goal is not only to raise technical performance but also to establish an AI platform that underpins national competitiveness across industry, research, and security.​

Architecture and Efficiency Innovations

K-EXAONE centers on an efficiency-driven strategy, reducing training and inference costs while scaling model capacity. Instead of simply increasing parameter counts, LG AI Research designed a structure that can deliver high performance with more efficient use of compute and memory resources. This design allows K-EXAONE to reach frontier scale without requiring the same level of infrastructure investment as some global competitors.​

Mixture-of-Experts with 236B Parameters

Earlier EXAONE models used dense architectures, but K-EXAONE adopts a Mixture-of-Experts (MoE) structure to maximize scalability and efficiency. The model consists of 128 experts and activates the top 8 plus one shared expert for each token, meaning 9 experts operate simultaneously at inference time. Although the total parameter count is around 236 billion, only about 10%—approximately 23 billion parameters—are active for any given token, significantly reducing resource usage while preserving representational power.​

Multi-Token Prediction for Faster Decoding

Conventional language models rely on autoregressive decoding, predicting one token at a time and feeding each output back into the model, which can slow inference. K-EXAONE integrates a Multi-Token Prediction (MTP) module that allows it to predict the next token and the subsequent token simultaneously, effectively drafting multiple future tokens. This self-speculative decoding improves throughput and yields roughly 1.5× faster inference compared to standard one-token-at-a-time decoding, enabling noticeably quicker responses.​

Advanced Tokenizer and SuperBPE

Tokenizer design is another key lever for efficiency: if each token encodes more information, fewer tokens are needed to represent the same text. Compared with earlier EXAONE models that used roughly 100,000 vocabulary items, K-EXAONE expands its vocabulary to around 150,000 tokens, improving coverage across languages and specialized domains like science, technology, mathematics, and code.​

The model also employs a SuperBPE strategy that groups common multi-word expressions into single “Superword Tokens,” such as converting “New York” into a single token like “New_York,” and these superwords make up about 20% of the tokenizer’s vocabulary. As a result, K-EXAONE’s tokenizer achieves around a 30% efficiency gain in bytes per token versus the previous EXAONE tokenizer, which translates into lower processing overhead for many real-world texts.​

Training Strategy: From Pre-Learning to Post-Learning

K-EXAONE’s training pipeline is designed to build knowledge, reasoning, and alignment in stages rather than relying on a single training phase. This multi-stage approach is crucial for achieving both strong raw performance and behavior that matches human expectations in practical use.​

Three-Stage Pre-Learning

Pre-learning for K-EXAONE unfolds in three stages that progressively add foundational knowledge, domain-specific expertise, and reasoning skills. The final stage incorporates synthetic data that includes explicit “thinking trajectories,” allowing the model to learn step-by-step reasoning rather than only mapping inputs to final answers. This method helps the model internalize logical flows and intermediate steps, which later strengthens reasoning behavior during post-learning.​

Extending Context to ~260K Tokens

A standout feature of K-EXAONE is its ability to handle very long contexts, supporting up to approximately 260,000 tokens by progressively extending its context window. In practice, the base model is first extended from about 8,000 tokens to roughly 33,000 tokens, and then further to a context window around 256,000–260,000 tokens using a two-stage context length extension procedure.​

Simply training on long-context data can harm performance on shorter inputs, so K-EXAONE’s training uses a rehearsal dataset of high-quality short-context examples alongside long-context data to preserve performance across both regimes. To verify that the extended context is actually usable, the team periodically runs Needle-In-A-Haystack (NIAH) tests, which check whether the model can reliably retrieve specific pieces of information embedded in long documents.​

Post-Learning: SFT, RL, and Preference Alignment

Post-learning refines how K-EXAONE uses its knowledge to follow instructions, solve problems, and interact with users. It is described as the stage where the model learns to apply its textbook-like knowledge to real-world tasks and user queries. This stage is organized into supervised fine-tuning, reinforcement learning, and preference alignment.​

(1) Supervised Fine-Tuning with Virtual Tools

Supervised Fine-Tuning (SFT) teaches the model to understand and respond to diverse instructions, with a particular focus on tool use as AI moves toward more agentic behavior. Because connecting to real-world tools and APIs at scale is expensive and complex, K-EXAONE uses a virtual tool environment to generate training data, simulating coding tools and other common tool invocation scenarios with verifiable success criteria for each task. Through this approach, the model learns hundreds of realistic and checkable tool-usage tasks, laying the groundwork for practical agent behavior in real environments.​

(2) Reinforcement Learning with AGAPO

Reinforcement learning (RL) is used to further enhance reasoning in domains such as mathematics, coding, science, and instruction following. K-EXAONE’s RL stage is driven by verifiable rewards—for example, assigning 1 point when generated code passes tests and 0 when it fails—so that feedback is objective and grounded in clear outcomes.​

Instead of relying on the GRPO (Group Relative Policy Optimization) algorithm, K-EXAONE introduces AGAPO (Asymmetric Sampling and Global Advantage Policy Optimization), a newly designed RL algorithm. While GRPO tends to discard samples where all responses are either entirely correct or entirely incorrect, AGAPO makes use of even all-incorrect cases by assigning small negative rewards and guiding the model away from those failure patterns, allowing it to learn more from mistakes and refine inference paths more effectively.​

(3) Preference Alignment with GrouPER

After boosting reasoning and task performance, K-EXAONE undergoes a preference learning phase to adjust its responses to human expectations around safety, clarity, and style. For this, it employs GrouPER (Group-wise SimPER), which extends the SimPER algorithm by comparing multiple responses for each prompt at once instead of evaluating them individually.​

For each prompt, several candidate responses are generated and then ranked using both rule-based criteria—such as whether the answer is safe and instruction-following—and rubric-based criteria, including creativity, logical coherence, and naturalness. Group-wise comparison helps stabilize learning by averaging out noise from individual samples, and the model is updated based on which responses receive higher preference scores within each group.​

Benchmark Results and Global Standing

Performance evaluation for K-EXAONE focuses on how effectively it can be applied in real-world scenarios, not just whether it gets answers right on a single test. Since each benchmark tends to target specific capabilities, a broad set of tests across domains is used to build a comprehensive performance profile.​

K-EXAONE has been evaluated on benchmarks covering knowledge (MMLU-Pro), mathematics (AIME 2025), coding (LiveCodeBench v6), agent tool usage (τ²-Bench), instruction following (IFBench), Korean language (KoBALT), multilingual capabilities (MMMLU), and safety (KGC-Safety). In these comparisons, K-EXAONE is tested against the EXAONE 4.0 32B model as well as top open-weight peers like gpt-oss-120b (GPT), Qwen3-235B-A22B-Thinking-2507 (Qwen3), and DeepSeek-V3.2.​

The model demonstrates strong competitiveness in knowledge, math, and coding, achieving excellent results on MMLU-Pro for academic knowledge and reasoning and surpassing GPT and Qwen3 on the AIME 2025 math benchmark. On LiveCodeBench v6, K-EXAONE shows improved coding performance over Qwen3 and DeepSeek, highlighting its suitability for real-world programming tasks. It also performs well on τ²-Bench for tool-usage scenarios that require multi-step decision-making, and it maintains solid scores on IFBench, KoBALT, and MMMLU, indicating stable instruction following, Korean fluency, and multilingual capability.​

In the domain of safety, K-EXAONE achieves top results on the KGC-Safety benchmark, demonstrating robust defenses against harmful prompts while aligning with Korean sociocultural norms and broader global ethical standards. This balance shows that the model can manage sensitive topics responsibly without sacrificing general-task performance, which is critical for deployment in public and enterprise settings.​

National Impact and Future Directions

K-EXAONE’s achievements are reflected in its ranking on the Artificial Analysis Intelligence Index, where it places 7th globally and 1st among Korean open-weight models, representing Korea’s sole entry in a ranking otherwise dominated by US and Chinese models. This high placement signals not just strong numerical scores but also Korea’s emergence as a serious player in the global foundation model landscape. It underscores the notion of “AI sovereignty,” where Korea can field a homegrown model that competes at the frontier level while reflecting local priorities.​

Technically, K-EXAONE has raised the bar for Korean AI by combining an MoE structure, multi-token prediction, and a highly efficient tokenizer with long-context support up to around 260,000 tokens. Through its multi-stage pre-learning and post-learning pipeline—covering supervised fine-tuning, reinforcement learning with AGAPO, and preference alignment with GrouPER—the model attains high accuracy and strong safety across a wide variety of tasks. Beyond benchmarks, K-EXAONE is positioned as a practical system for industry and research, aiming to solve real problems, support more informed decision-making, and drive innovation as Korea’s leading national AI representative in the global market.