CtrlK
Naver AI refers to the artificial intelligence research and products developed by Naver Corporation, South Korea's largest internet company. Naver's AI efforts center on the HyperCLOVA family of large language models, the CLOVA AI research division, and the CLOVA Studio platform for enterprise AI deployment. Naver has positioned itself as a leader in Korean-language AI, building models trained on vastly more Korean data than any Western counterpart, and has expanded its AI capabilities into search, e-commerce, cloud services, robotics, and healthcare.
Naver Corporation is a South Korean internet conglomerate headquartered in Seongnam, Gyeonggi Province. Founded by Lee Hae-jin in 1999, the company operates the Naver search engine, which has been South Korea's dominant search portal for over two decades. The company originated as the "Web Glider" project within Samsung SDS in 1997 before spinning off as an independent entity in 1999.
Naver was the first Korean web provider to develop its own search engine. In August 2000, it launched its "comprehensive search" service, which aggregates results from blogs, websites, images, and web communities into a single page. This approach to organizing search results became a defining feature of Korea's internet experience and helped Naver maintain dominance over Google in the Korean market for many years.
The company's current business spans search, advertising, e-commerce, digital content (including Naver Webtoon), fintech, cloud computing, and AI. Key subsidiaries and affiliates include Naver Cloud, Naver Labs, Naver Webtoon, Snow, and Works Mobile. As of early 2026, Naver's market capitalization stands at approximately $22 to $29 billion USD, and the company reported revenue of 12.04 trillion won (roughly $8.3 billion USD) in 2025.
Naver's second headquarters building, called "1784" (named after its address at 178-4 Jeongja-dong and the year the Industrial Revolution began), opened as a 28-story, robot-friendly smart building. Over 100 autonomous service robots operate inside, delivering items to employees via a dedicated robot-only elevator system called Roboport. The building serves as a testbed for Naver Labs' robotics and digital twin technologies.
Naver introduced its artificial intelligence platform CLOVA (short for "Cloud Virtual Assistant") in March 2017. Since then, CLOVA and Naver's AI Lab have invested continuously in large-scale AI research and development.
The CLOVA AI Research team is responsible for fundamental AI technologies based on machine learning and deep learning that power the Naver and LINE AI platforms. Naver's research spans computer vision, natural language processing, speech recognition, recommendation systems, and pattern recognition. The team publishes regularly at top-tier AI conferences including NeurIPS, ICML, CVPR, ECCV, EMNLP, and ACL.
Jung-Woo Ha has led Clova AI Research and served as the head of Naver AI Lab since joining Naver Labs as a staff research scientist in 2015. Under his leadership, the team grew from a small group into one of Asia's most prolific AI research organizations.
CLOVA's earliest consumer-facing product was a virtual assistant that powered smart speakers. In 2017, Naver and LINE Corporation launched the CLOVA Wave smart speaker in Japan and Korea, followed by the CLOVA Friends speakers (featuring LINE's popular Brown and Sally characters) in December 2017. These devices supported voice commands for reading messages, creating memos, online shopping, calling taxis, controlling home appliances, and streaming music.
HyperCLOVA was unveiled at the "Naver AI NOW" conference on May 25, 2021, as the first hyperscale AI language model independently developed by a Korean company. The model was designed as a Korean-centric variant of the GPT-3 architecture.
The flagship HyperCLOVA model contained 82 billion parameters and was trained on a Korean-centric corpus of 560 billion tokens. Naver also developed mid-size variants with 39 billion parameters, which the team noted could be more practical for real-world applications. At the time of its release, Naver stated that HyperCLOVA had been trained on 6,500 times more Korean data than GPT-3, whose training corpus was overwhelmingly English.
The model used a Korean-specific tokenizer optimized for the Korean language's agglutinative morphology, allowing for more efficient processing of Korean text compared to tokenizers designed primarily for English.
The research behind HyperCLOVA was published as a long paper at EMNLP 2021, titled "What Changes Can Large-scale Language Models Bring? Intensive Study on HyperCLOVA: Billions-scale Korean Generative Pretrained Transformers." Led by Boseop Kim and a team of 37 authors, the paper demonstrated state-of-the-art in-context zero-shot and few-shot learning performances on various downstream tasks in Korean. It also introduced HyperCLOVA Studio, an interactive prompt engineering interface that foreshadowed the later CLOVA Studio product.
HyperCLOVA was significant as one of the first non-English hyperscale language models, demonstrating that language models trained primarily on a single non-English language could match or exceed the performance of English-centric models on tasks in that language. It established Naver as a serious competitor in the global LLM landscape and set the stage for subsequent Korean AI development.
Naver unveiled HyperCLOVA X on August 24, 2023, as a substantially upgraded successor to the original HyperCLOVA. While HyperCLOVA was primarily Korean-focused, HyperCLOVA X was designed from the ground up as a bilingual Korean-English model with strong multilingual capabilities.
HyperCLOVA X is built on a transformer decoder architecture with several key modifications: rotary position embeddings for increased context length, pre-normalization for training stability, and grouped-query attention for improved efficiency. The model family consists of two sizes: a larger model (HCX-L) and a smaller model (HCX-S), though Naver has not publicly disclosed the exact parameter counts.
The model uses a morpheme-aware byte-level byte pair encoding tokenizer with a vocabulary size of 100,000 tokens. This tokenizer achieves superior compression efficiency compared to competing tokenizers across multiple languages, which is particularly important for Korean, where standard English-optimized tokenizers often produce excessively long token sequences.
The pretraining data consists of roughly equal portions of Korean, multilingual content (predominantly English but also including Japanese, German, French, and other languages), and code. Training was conducted with a context length of 4,096 tokens for 90% of the process, with the final 10% extending to 32,768 tokens. After pretraining, the model underwent supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) using high-quality human-annotated datasets.
One of HyperCLOVA X's standout features is its cross-lingual transfer ability. Instruction tuning in one language leads to the emergence of instruction-following capabilities in the other. The model demonstrates strong proficiency in Korean, English, Japanese, and Chinese, and shows unexpectedly strong performance on languages that were underrepresented in training data, including Arabic, Hindi, Thai, Urdu, Vietnamese, and Tagalog.
The HyperCLOVA X technical report, published on arXiv in April 2024 with over 395 contributors, presents evaluation results across comprehensive reasoning, knowledge, commonsense, factuality, coding, math, chatting, instruction-following, and harmlessness benchmarks. Notable scores include 81.65% on GSM8K (math reasoning), 55.49% on HumanEval (code generation), and 72.07% on comprehensive Korean language tests.
In 2025, Naver released the HyperCLOVA X SEED series, its first open-source model lineup. SEED (which stands for the idea of planting seeds in the AI ecosystem) makes Naver's Korean-optimized language technology available for commercial use by businesses, researchers, and developers worldwide. All SEED models are available on Hugging Face under the naver-hyperclovax organization.
| Model | Parameters | Type | Key Features |
|---|---|---|---|
| SEED Text Instruct 0.5B | 0.6B | Text | Lightweight text generation |
| SEED Text Instruct 1.5B | 2B | Text | Instruction-following, 16K context |
| SEED Vision Instruct 3B | 3.6B (3.2B LLM + 0.43B vision) | Vision-Language | First Korean open-source vision-language model |
| SEED Omni 8B | 11B | Multimodal | Unified text, vision, and speech; any-to-any generation |
| SEED Think 14B | ~14B | Reasoning | Pruning and knowledge distillation with RL-based reasoning |
| SEED Think 32B | 33B | Vision-Language Reasoning | 128K context, controllable thinking mode, multimodal reasoning |
The SEED Think models use two key technologies: pruning combined with knowledge distillation to achieve compactness while maintaining high performance, and a reinforcement learning pipeline to maximize reasoning ability. The SEED Omni 8B model is a unified any-to-any model that processes and generates text, images, and audio through a single architecture.
CLOVA Studio is Naver Cloud Platform's enterprise AI service built on HyperCLOVA technology. It functions as both an API platform and an LLMOps (Large Language Model Operations) platform for managing the full lifecycle of LLM-based applications.
The platform provides RESTful APIs for sentence generation, model tuning, an explorer interface, a router for model selection, and a skill trainer for custom capabilities. Additional specialized tools include token calculation, text summarization, paragraph division, embedding generation, and translation.
CLOVA Studio uses an API key authentication system, with each Naver Cloud Platform account able to create up to 10 keys. The platform supports the full development workflow: design, development, deployment, and ongoing operation of LLM-based applications from a single interface.
As of 2025, more than 2,000 companies and institutions use CLOVA Studio for their AI applications. Enterprise customers span industries including finance, healthcare, retail, education, and government services.
CLOVA X is Naver's consumer-facing chatbot application, launched in beta on August 24, 2023, alongside the HyperCLOVA X announcement. It serves as Naver's answer to ChatGPT for Korean users.
The chatbot supports both Korean and English and is designed to boost work productivity through features like document summarization, language translation, email drafting, blog post composition, story creation, and code generation. CLOVA X can also enhance web searches, online shopping experiences, and browsing activities.
Naver has continued developing CLOVA X's multimodal capabilities. HyperCLOVA X Vision, introduced in 2024, added the ability to analyze and generate content across text, images, video, and audio modalities.
Naver has integrated HyperCLOVA X into its core search engine through several features, representing its version of AI-enhanced search similar to Google's AI Overviews or Perplexity AI.
Naver launched "Cue:" as an AI-powered search chatbot in September 2023 for beta testing, incorporating it into Naver's search engine in November 2023. Cue: provided concise, AI-generated answers to user queries. The service was later discontinued in March 2025 as Naver shifted its AI search strategy toward AI Briefing.
AI Briefing, officially launched on March 26, 2025, became Naver's primary AI search feature. It functions as an intelligent assistant integrated directly into search results, understanding user intent and context to provide concise information summaries while citing original sources for verification.
By December 2025, AI Briefing appeared in more than 20% of all search queries on Naver. The feature expanded throughout the second half of 2025 into entertainment, stock market searches, location-based services ("Place AI Briefing"), and other specialized domains.
In 2026, Naver announced plans for an "AI Tab" within search results, scheduled for the first half of the year. This feature aims to act as a smart assistant capable of handling complex, context-aware tasks where users can search for a topic and complete related actions (restaurant reservations, trip planning, purchases) within a single conversational flow. This represents Naver's push into agentic AI for search.
CLOVA CareCall is one of Naver's most socially impactful AI applications. Launched as a welfare check service, it uses HyperCLOVA's conversational AI to make regular phone calls to elderly people living alone, checking on their well-being through natural conversation.
The system features a "Remember" function that retains conversational context across calls. If a user mentions plans to visit a hospital during one call, CareCall will follow up about the visit in the next conversation and ask about any prescribed medications.
As of 2025, CLOVA CareCall has been implemented in approximately 150 municipalities across South Korea, including Seoul, Busan, Daegu, and Gwangju, serving an estimated 50,000 older adults. Research published by Naver Cloud and Yonsei University found a 44.2% decrease in solitary deaths in localities where the program operates. Surveys conducted by local governments showed that 90% of senior users felt consolation from the calls, and 95% wanted to continue using the service.
Naver's AI strategy is closely tied to its relationship with LINE Corporation and the Japanese market. LINE, a messaging app with hundreds of millions of users across Japan and Southeast Asia, was originally developed by Naver's Japanese subsidiary.
On October 1, 2023, Z Holdings (which owned LINE Corporation and Yahoo! Japan) merged with its subsidiaries to form LY Corporation. Naver Corporation and SoftBank Corp. each hold 50% stakes in A Holdings Corp., which holds a majority stake in the merged entity. This gave Naver significant influence over one of Japan's largest internet companies.
However, in October 2023, a security incident resulted in the leak of 520,000 records of personal data from LINE users. In March 2024, the Japanese government ordered LY Corporation to sever technical system ties with Naver, creating tension in the partnership.
Naver has worked to bring HyperCLOVA X to the Japanese market through B2B services integrated into LINE Works and Naver Works productivity tools. HyperCLOVA X demonstrates strong proficiency in Japanese, benefiting from the model's multilingual training data and cross-lingual transfer capabilities. LINE AiCall, an AI-powered restaurant booking service, uses HyperCLOVA technology to facilitate natural voice conversations.
Naver has invested heavily in AI compute infrastructure to train and serve its models.
In early 2026, Naver Cloud completed a 4,000-GPU cluster using NVIDIA B200 GPUs, the largest AI computing deployment in South Korea. The company reported that training a 72-billion-parameter model that previously took around 18 months on its older A100-based infrastructure (2,048 GPUs) can now be completed in roughly six weeks, representing a twelvefold improvement in training efficiency.
Naver has announced plans to invest over 1 trillion won (approximately $700 million) in GPU infrastructure. The company also operates one of Asia's largest single-company data centers, the GAK Sejong facility, a 294,000-square-meter complex with capacity for up to 600,000 rack units and 65 exabytes of storage.
On March 18, 2026, Naver CEO Choi Soo-yeon and AMD CEO Lisa Su signed a memorandum of understanding on AI ecosystem expansion and next-generation infrastructure cooperation. Under the agreement, Naver Cloud is expanding its use of AMD EPYC processors to include the upcoming 6th-generation AMD EPYC "Venice" processor, and AMD is providing access to next-generation AMD Instinct MI455X GPUs. The partnership aims to build a high-performance GPU computing environment optimized for HyperCLOVA X and establish a more diversified AI infrastructure ecosystem beyond NVIDIA.
Naver has embraced the concept of "sovereign AI," building AI systems designed to align with each nation's unique language, culture, and data sovereignty requirements.
In August 2025, South Korea's Ministry of Science and ICT selected five consortia to compete for sovereign AI development, with $381 million in government funding. Naver was selected as one of the five lead organizations, alongside SK Telecom, LG Group, NCSoft, and Upstage.
Naver has been particularly active in the Middle East, establishing Naver Arabia as its regional base. Key initiatives include:
Naver competes with several major Korean technology companies developing their own AI models and services.
| Company | AI Model(s) | Notable Details |
|---|---|---|
| Naver | HyperCLOVA X, SEED series | Trained on 6,500x more Korean data than GPT-3; open-source SEED lineup |
| SK Telecom | A.X series (A.X 3.1) | 34B parameters; built from scratch with proprietary technology (2024) |
| Kakao | Kanana (Nano, Essence, Flag) | Nano 2.1B, Essence 9.8B, Flag 32.5B parameters |
| Upstage | Solar Pro 2 | 31B parameters; first Korean model recognized by Artificial Analysis as frontier |
| LG | EXAONE | LG AI Research's multilingual model |
| NCSoft | VARCO | Gaming-focused AI with broader LLM capabilities |
| Samsung | Samsung Gauss | On-device AI models for Galaxy products |
Naver's primary competitive advantages are its massive Korean-language training dataset, its integrated ecosystem (search, e-commerce, content, cloud), and its established relationship with Korean enterprise customers through CLOVA Studio. However, the Korean AI market faces pressure from global competitors, particularly OpenAI's ChatGPT and Google's Gemini, which have gained significant traction among Korean users.
Choi Soo-yeon has served as CEO of Naver since 2022. A graduate of Seoul National University and Harvard Law School, she became the company's youngest-ever CEO at age 40. Under her leadership, Naver surpassed $7.5 billion in annual revenue for the first time in 2024. Choi has focused on integrating AI into Naver's core businesses, stating that "AI should not be a standalone service but must be naturally embedded into our core businesses, such as search, advertising, commerce and content."
In 2025, Naver founder Lee Hae-jin returned as chairman of the board to drive the company's sovereign AI strategy. The same year, CEO Choi was appointed to the United Nations Global Compact Board of Directors for her leadership in AI governance and recognized by Fortune Asia as one of the Most Powerful Women in Asia.
| Year | Milestone |
|---|---|
| 1999 | Naver Corporation founded by Lee Hae-jin |
| 2000 | Launch of Naver's comprehensive search service |
| 2017 | CLOVA AI platform introduced; Wave and Friends smart speakers launched |
| 2021 | HyperCLOVA unveiled (82B parameters, 560B Korean tokens); CLOVA CareCall launched |
| 2023 | HyperCLOVA X announced (bilingual Korean-English); CLOVA X chatbot and Cue: search launched; LY Corporation formed from LINE-Yahoo Japan merger |
| 2024 | HyperCLOVA X technical report published on arXiv; HyperCLOVA X Vision introduced; Japanese government orders LINE-Naver system separation |
| 2025 | HyperCLOVA X SEED open-source models released; AI Briefing launched for Naver Search; Selected for South Korean sovereign AI program; Naver Arabia established; CLOVA CareCall reaches 150 municipalities |
| 2026 | 4,000 NVIDIA B200 GPU cluster completed; AMD infrastructure partnership signed; AI Tab for agentic search announced |