Naver

corporate

korea · Founded 1999 · IPO 2002 (KRX: 035420) · Market cap: $19B (KRX: 035420) as of 2026-04

Naver (KRX: 035420), South Korea's leading internet conglomerate, is a global pioneer in "Sovereign AI" through its research arms: NAVER AI Lab (Seoul), NAVER Labs Europe (Grenoble), and NAVER Cloud. R&D spending hit a record 2.22 trillion KRW (~$1.67B) in 2025 (18.5% of revenue), and the company completed South Korea's largest AI cluster with 4,000 NVIDIA B200 GPUs in January 2026.

The flagship HyperCLOVA X model family evolved from the original HyperCLOVA (2021, 82B) through the HyperCLOVA X technical report (2024) to the current open-weight SEED Think (32B, Dec 2025) reasoning VLM and SEED Omni (8B) any-to-any multimodal model. HyperCLOVA X THINK (June 2025) was pre-trained on ~6T Korean/English tokens with Peri-LN Transformer and muP scaling.

NAVER Labs Europe has produced landmark 3D vision work: DUSt3R (CVPR 2024, 7K+ GitHub stars) reformulated 3D reconstruction as pointmap regression, and MASt3R extended it with dense matching. The former Clova AI lab (now integrated into NAVER AI Lab) produced widely-used tools including Donut (OCR-free document understanding, ECCV 2022), CRAFT (text detection, CVPR 2019), StarGAN (CVPR 2018), and SPLADE sparse retrieval.

Website Wikipedia HuggingFace GitHub
open-weightmultilingualmultimodalvisionretrieval

People

Outputs (11)

HyperCLOVA X SEED Omni

model

First any-to-any omnimodal model in the HyperCLOVA X family. Supports text, audio, and vision as both inputs and outputs via unified next-token predic…

HyperCLOVA X SEED Think

model

Vision-language reasoning model with chain-of-thought and agentic capabilities. 32B parameters, 128K context. Supports text, image, and video input. A…

HyperCLOVA X THINK

model

<p>First reasoning-focused LLM in the HyperCLOVA X family. Exact parameter count not disclosed (analysts estimate ~32B based on SEED distillation). Pr…

MASt3R

paper

Extends DUSt3R with dense local features and matching loss. Beats prior SOTA by 30% in VCRE AUC on Map-free localization. 2.8K+ GitHub stars. From NAV…

HyperCLOVA X

model

Foundation model family optimized for Korean language and culture with competitive bilingual capabilities. Reported as 204B+ parameters (exact count n…

DUSt3R

paper

Landmark paper from NAVER Labs Europe reformulating 3D reconstruction as pointmap regression. Recovers 3D models, depth, correspondences, and camera p…

Donut

paper

OCR-free Document Understanding Transformer. End-to-end model for visual document understanding that eliminates OCR dependency, error propagation, and…

HyperCLOVA

model

Korean GPT-3 variant trained on 560B tokens. The arxiv paper covers the 82B model; press reports indicate a 204B variant was also trained but not publ…

SPLADE

paper

Sparse Lexical and Expansion Model for first-stage ranking. Combines advantages of dense embeddings with inverted index efficiency using explicit spar…

CRAFT

paper

Character Region Awareness for Text Detection. Scene text detection using character-level region awareness, excelling at curved and irregularly-shaped…

StarGAN

paper

Unified GAN for multi-domain image-to-image translation, eliminating the need for separate models per domain pair. From Clova AI Research. Published a…