LG AI Research Unveils 750-Billion-Parameter K-EXAONE 2.0
LG AI Research has unveiled K-EXAONE 2.0, a large-scale artificial intelligence (AI) model with 750 billion parameters, more than tripling the size of its predecessor while adopting a commercially permissive open-source license.
The institute said on July 31 that it had released K-EXAONE 2.0 on Hugging Face as the second model developed under South Korea’s Independent AI Foundation Model Project, an initiative overseen by the Ministry of Science and ICT.
The model does not activate its entire network for every request. Instead, it employs a Mixture of Experts (MoE) architecture that selectively engages only the components needed to answer a query. The system contains 256 specialized processing modules and dynamically selects eight experts for each token when generating responses. While the model contains 750 billion parameters in total, only 37 billion parameters are activated during inference. The approach is designed to preserve the knowledge capacity of a large model while reducing computational requirements and operating costs.
K-EXAONE 2.0 supports a context window of up to 262,144 tokens and expands language coverage to 10 languages, including Korean, English, Japanese, and Chinese. LG AI Research also replaced its previous proprietary license with the Apache 2.0 license, allowing users to download, modify, and deploy the model commercially.
According to LG AI Research, the model achieved an average score of 70.1 across 24 benchmarks spanning nine evaluation categories, compared with 63.3 for the first-generation model. In the OpenAI-MRCR long-context retrieval benchmark, K-EXAONE 2.0 scored 94.4, exceeding the 93.0 achieved by Qwen 3.5 and the 92.9 recorded by DeepSeek V4 Pro Max.
Within South Korea, one of the most direct comparison points is SK Telecom’s A.X K2 model, unveiled on July 29. A.X K2 is also an MoE model, with 688 billion total parameters and 33 billion active parameters. Like K-EXAONE 2.0, it supports a maximum context length of 262,144 tokens and is released under the Apache 2.0 license.
The announcement did not include direct comparisons with leading commercial AI models such as OpenAI’s GPT series, Anthropic’s Claude, or Google’s Gemini. As a result, the published benchmark data alone does not provide a clear indication of how K-EXAONE 2.0 compares with the highest-performing proprietary global models.
Parameter count alone is also not a direct measure of real-world performance. User experience depends not only on reasoning capability but also on factors such as response speed, operating costs, integration with search and coding tools, multimodal capabilities, and service optimization.
The significance of the release lies in the fact that South Korean researchers have built a 750-billion-parameter-class model, including its training and inference infrastructure, and made it available under terms that allow enterprises to maintain control over deployment and customization.
The model’s competitive standing is expected to become clearer once results emerge from government-led evaluations conducted under standardized conditions, independent third-party benchmarks, and real-world deployment scenarios.