Naver’s Tailored LLM and Multimodal AI Search: How AI Tab Is Redefining the Search-to-Action Journey

From 27 Years of Search to an AI-Native Experience

Naver is refactoring 27 years of search infrastructure, logs, and UGC from Blog, Café, Shopping, and Place into an AI-native stack that connects a query directly to buying, booking, or visiting.[1] This is end-to-end orchestration over a mature content and commerce ecosystem, not just retrieval.[1]

AI Tab is the flagship interface for this shift:

• Generative, conversational surface on mobile and PC
• Anchored at the search bar and Green Dot entry points[5]
• Default “search brain” for text, voice, image, and location queries[5]

Impact at scale:[5]

• Naver main site: ~50M daily visitors
• AI Tab beta: 4M+ cumulative users
• Product and location card CTRs: 20%+
• Heavy users (11+ uses):
  • 2.7× more product clicks
  • 2× more location clicks vs. one-time users

This reflects a shift from benchmark scores to “practical AI” tuned for hundreds of millions of daily queries.[3][4] For natural-language queries like “restaurants with convenient parking,” the system must interpret nuanced reviews, not just tags like “parking available.”[3][4]

Key takeaway: Naver is rewiring search as a continuous journey from intent → understanding → concrete action at massive scale, not building a generic chatbot.[1][3]

Inside Naver’s Product-Native LLM and Harness Engineering

Naver’s product-native LLM is a lightweight model based on HyperCLOVA X, optimized for search, shopping, and reservation flows rather than generic NLP tasks.[1][4]

Core design choices:

• Domain-tuned data:
  • Trained on Naver logs and service usage
  • High-quality content from search, shopping, places, lifestyle via filters and pipelines[3]
• Architecture:
  • Transformer foundation with web-scale pretraining[6][7]
  • Mixture-of-Experts (MoE): activates only part of the model per request for faster, cheaper inference under heavy traffic[1][3]
• Training focus:

  • 2× reinforcement learning resources vs. HyperCLOVA X to optimize real-world performance[1]

  • Clarity-focused RL: model prefers clarifying questions over confident guesses, cutting hallucinations by up to 30%[1][4]

Performance vs. HyperCLOVA X:[1][4]

• 2× faster responses

• Up to 3× lower operating cost
• Significantly fewer hallucinations

The LLM is trained with agentic AI capabilities to act across Naver services like Maps and reservations:[4][5]

• Example flow: “Find a kid-friendly brunch place near Gangnam with easy parking” → refine via chat → transition into real-time booking in the same guided experience.[5]

Beneath the model, a harness engineering layer turns the LLM into a robust service:[3][8]

• Prompt and tool orchestration
• Evaluation frameworks and routing logic
• Safety filters and observability between the raw model and live endpoints[8]

Research on efficient inference, such as speculative decoding in EdgeLLM (up to 9.3× faster token generation on constrained devices), underscores why runtime behavior is as critical as parameter count.[7] Naver’s MoE, small models, and orchestration adopt this “do more with less, fast” philosophy at cloud scale.[1][3][7]

Implementation lens: For ML/product teams, this translates to domain-specific data pipelines, MoE architectures, RL for interaction quality, and a strong orchestration tier for evaluation, routing, and tooling.[1][3][8]

Multimodal AI Search: OmniSearch, Smart Lens, and AI Tab

Naver defines multimodal search as jointly processing text and images so a single model, OmniSearch, can align keywords and pictures with relevant content across Blog, Café, Shopping, Knowledge iN, and News.[2] This “multimodal document search” ranks documents by how well they match both textual and visual signals.[2]

Example: sneakers[2]

• User uploads or captures an image without knowing the product name
• System returns matched items, reviews, and styling ideas
• Results appear as a “reviews and styles searched by image” block for quick decision-making

Multimodal document search uses smart thumbnail technology:[2][3]

• Automatically crops images to the most relevant region
• Improves scanability and click-through by emphasizing the clearest visual cue
• This visual understanding increasingly feeds into AI Tab

Smart Lens, Naver’s multimodal search tool, is placed beside the AI Tab button in the Green Dot interface, with music search integrated as well.[5] Voice, images, and text now converge into a single conversational, action-oriented surface, shrinking the gap between “what is this?” and “book/buy this now.”[2][5]

As models learn to jointly interpret photos, maps, menus, and long-tail reviews, AI Tab can act as a proactive agent for local discovery:[3]

• Instead of only answering “Which café?”, it can offer:
  • A few candidates
  • Live wait times
  • Reservation options aligned with user preferences

Key takeaway: Multimodality grounds language in real-world objects, places, and behaviors so search becomes context-rich and immediately actionable.[2][3][5]

Conclusion: A Playbook for Service-First Generative AI

Naver’s AI stack—product-native LLM, harness engineering, and multimodal models—shows how to build generative AI around service quality rather than model size.[1][3] By tuning to Korean user behavior, deep-linking into commerce and local services, and prioritizing efficiency, Naver converts complex intents into real-world actions at massive scale.[3][5]

Practical lesson: Domain-specific tuning, efficiency-first infrastructure, and tight integration with real product surfaces often beat chasing generic, maximal models.[1][3][8]