Upstage will provide the Daum portal’s "AI Summary" service, operated by its subsidiary AXZ, based on its Solar large language model (LLM) and FuriosaAI’s Renegade (RNGD) neural processing unit (NPU).

Upstage CEO Kim Sung-hoon (김성훈) said at a meeting on July 15 with AXZ CEO Lee Kun-soo (이건수) and FuriosaAI CEO Baek Jun-ho (백준호) that "this is the first case linking the entire AI domain, from infrastructure to models to services, with domestic technology."

The three CEOs held a discussion that day at AXZ’s headquarters in Pangyo, Seongnam, Gyeonggi Province, and discussed cases of cooperation among the three companies and future plans.

Daum AI Summary is a function in which an LLM directly analyses web documents and organises comprehensive answers to search results. When users enter a keyword in Daum’s search bar, it provides key summaries along with supporting grounds. When content changes, the AI automatically reflects it so users can always check the latest information.

Lee said "20 percent of the total query volume is being handled by AI Summary" and "we are preparing to expand beyond simple summaries into verticals such as shopping and restaurant searches, and also to distribute a specialised 'one agent per person' to users through Daum." Daum began a beta AI Summary service on July 1. AXZ plans to broaden the areas of application within the year and distribute an official version.

He added that "we advanced harness engineering to control hallucinations" and explained that "we are delivering highly relevant, up-to-date information to the Solar LLM through hybrid search that combines keyword search with vector-based search."

On computing resources for AI Summary processing, Baek said "we are processing 500 million tokens a day with 3 nodes" and "we are delivering performance comparable to the H200 by building hardware and software through original designs, including development of a compiler that can map AI models well to accelerators and an engine for optimal serving." He added that introducing FuriosaAI’s NPU reduces inference costs by more than 50 percent.

Baek stressed that "considering the international situation, from a strategic perspective it is also necessary to combine sovereign AI semiconductors with models and services."

Kim cited Solar’s competitiveness as global frontier-level performance secured through in-house small model design and Korean-language specialisation capabilities. He said "we will further advance the performance of domestic LLMs that can stand up to overseas big tech models" and "we ask for a lot of interest in Upstage’s next-generation model, which will be unveiled soon, through independent AI foundation model projects and others."