LG AI Research, the corporate research arm of LG Group, released K-EXAONE 2.0 on July 31, 2026, making it available on Hugging Face under the Apache 2.0 license. <cite index="1-1">The model is a 750-billion-parameter artificial intelligence (AI) foundation model developed under South Korea's Sovereign AI Foundation Model Project.</cite> <cite index="1-5">LG describes K-EXAONE 2.0 as the largest AI foundation model developed in South Korea to date.</cite>
Architecture and Scale
<cite index="20-8">K-EXAONE 2.0 uses a hybrid-attention Mixture-of-Experts (MoE) architecture with 750 billion total parameters and approximately 37 billion active parameters per token.</cite> Rather than activating its full parameter count on every query, <cite index="3-5,3-6">the model employs a 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.</cite> <cite index="24-3">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.</cite>
<cite index="23-1">K-EXAONE 2.0 was scaled to more than three times the size of its predecessor through upcycling, followed by continual pretraining, difficulty-focused mid-training, and post-training.</cite> <cite index="1-8">This is more than three times the size of the first K-EXAONE model, which had 236 billion total parameters and 23 billion active parameters.</cite>
Benchmark Performance
<cite index="24-5">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.</cite> <cite index="8-16">Performance on three key coding and agentic coding benchmarks improved by about 30 percent.</cite>
Long-context retrieval was a particular area of strength. <cite index="24-6">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.</cite> <cite index="26-16">Those scores also exceeded Chinese Zhipu AI's GLM-5.1, which recorded 71.5 in English and 83.6 in Korean.</cite> On safety benchmarks, <cite index="20-1">K-EXAONE 2.0 achieved an average safety score of 94.6 across two benchmarks.</cite> LG's own technical report, however, <cite index="20-13">acknowledges that the model does not lead every test and still trails some competitors in certain areas.</cite>
Government Mandate and Licensing Shift
<cite index="6-4">K-EXAONE 2.0 is the second model developed by LG AI Research under a government-led program aimed at building an independent Korean foundation model and reducing reliance on foreign AI systems.</cite> <cite index="2-5">LG AI Research adopted the Apache 2.0 license for K-EXAONE 2.0, allowing unrestricted commercial use</cite>—a departure from the proprietary license applied to its predecessor. <cite index="7-11">LG AI Research also said it plans to unveil an industry-specialized AI model the following week.</cite>
Market Context: The Open-Weight Shift
The release arrives as open-weight models from Asia are measurably reshaping inference economics. <cite index="14-6,14-7">The share of tokens used by U.S. companies on Chinese AI models via OpenRouter—a platform that enables developers to access a range of AI models—has sat above 30% each week since February 8, 2026, with that figure rising as high as 46%; the average across the previous 12 months was just 11%, falling to 4.5% in the first half of 2025.</cite> <cite index="11-4">Open-source Chinese models are consistently 60% to 90% cheaper than the leading offerings from Anthropic and OpenAI, according to OpenRouter.</cite>
<cite index="10-4">OpenRouter's State of AI report identifies what researchers term a "Summer Inflexion" in mid-2025, when the open-source AI landscape shifted from near-monopoly to pluralistic competition.</cite> <cite index="13-11">Alibaba's Qwen family has spawned over 180,000 derivative models globally, making it one of the most adopted open-source large language model (LLM) families.</cite> K-EXAONE 2.0 positions South Korea within this competitive dynamic, offering an open-weight frontier-class alternative built independently of both U.S. and Chinese supply chains.
Woohyung Lim, co-head of LG AI Research, framed the release in sovereign terms: <cite index="2-10,2-11">"The significance of K-EXAONE 2.0 is not simply that it is a large-parameter model, but that Korean researchers independently completed the entire development process. It means we have secured the capability to compete with global frontier models at the same scale, which is the core objective of the National AI Foundation Model project."</cite>