HyperCLOVA X
NAVER's Korean-first LLM family, sold through CLOVA Studio
Overview
HyperCLOVA X is NAVER's large language model family, built for Korean language and culture and developed by NAVER Cloud. It ships as THINK (reasoning), DASH (fast and light) and open-weight SEED models from 0.5B to 32B, and is sold to businesses through CLOVA Studio after NAVER closed its consumer CLOVA X chatbot in April 2026.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- SEED 32B Think and 14B Think reasoning models
- SEED 8B Omni any-to-any model
- SEED 3B Vision, 1.5B and 0.5B
- "Powered by HyperCLOVA X" display required; >10M MAU needs a licence
- HCX-007, HCX-005 and HCX-DASH-002 models
- Input and output tokens billed separately
- Tuning, reranker and RAG reasoning tools
- SFT tuning for company-specific models
- Dedicated infrastructure
- Dedicated Neurocloud hardware with a GPU cluster
- HyperCLOVA X training and operation tools
Use cases
What you can produce with HyperCLOVA X
- Korean-language text generation, summaries and rewriting through the CLOVA Studio API
- Step-by-step reasoning over text and images with HyperCLOVA X THINK
- Company-specific models tuned on proprietary data via CLOVA Studio tuning or Exclusive
- Self-hosted Korean reasoning with the open-weight SEED 32B Think model
- Text, image and audio understanding and generation with the open SEED 8B Omni model
- Lightweight on-device Korean chat with the SEED 0.5B and 1.5B models
- On-premises HyperCLOVA X deployment through Neurocloud
ASEAN Perspective
HyperCLOVA X in Southeast Asia
A leading example of APAC sovereign AI — strongest for Korean-language and Korea-resident workloads, and relevant to ASEAN teams building for the Korean market or weighing regional, non-US model options. An open SEED variant is downloadable for self-hosting.
HyperCLOVA X is the default choice when Korean language quality and Korean data residency both matter. NAVER's own benchmarks show strong Korean results, the tokenizer is tuned for Korean text, and the SEED line gives self-hosters real options up to a 32B reasoning model and an 8B any-to-any model.
The caveats are commercial. The consumer CLOVA X chatbot closed in April 2026, CLOVA Studio rates sit behind NAVER Cloud's console rather than a readable price list, and the SEED licences add a 10-million-user ceiling, mandatory "Powered by HyperCLOVA X" branding and Seoul arbitration. Teams outside Korea should read the licence before building.
What we found
HyperCLOVA X is the large language model family NAVER unveiled in August 2023 and now develops through its cloud subsidiary, NAVER Cloud. Its selling point is Korean. NAVER says it was trained on 6,500 times more Korean data than GPT-4 and uses a Korean-optimised tokenizer that runs up to twice as fast at lower cost on Korean text; both are company figures. The line has three branches: THINK for step-by-step reasoning across text and images, DASH for lighter and faster work, and the open-weight SEED models.
The consumer side has gone. In February 2026 NAVER announced that its CLOVA X chatbot and Cue: AI search assistant would close on 9 April 2026, and the technology now sits inside AI Briefing and the AI tab in NAVER search. Citing a government survey, Newstomato reported that 2.0% of Korean generative-AI users had used CLOVA X, against 41.8% for ChatGPT. Earlier versions of this page dated a flagship refresh, to roughly 40% of the old parameter count and more than 50% cheaper to run, to February 2026; NAVER announced it in February 2025.
The open models are real but conditional. SEED 0.5B, 1.5B and 3B Vision arrived in April 2025 and SEED 14B Think in July 2025. In December 2025 NAVER Cloud released a 32B Think reasoning model and an 8B Omni any-to-any model as the first output of its entry in the government's sovereign foundation-model programme; in January 2026 the science ministry dropped that consortium after its first evaluation, citing originality criteria. Each SEED model has its own licence: services above 10 million monthly users need NAVER's permission, products must display "Powered by HyperCLOVA X", derived models must start with that name, and disputes go to arbitration in Seoul.
Paid access is CLOVA Studio on NAVER Cloud, billed per 1,000 tokens with input and output charged separately, plus CLOVA Studio Exclusive for dedicated tuning and Neurocloud for on-premises deployment. The rates load inside NAVER Cloud's pages by script and could not be read for this review.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- Trained on vastly more Korean data than GPT-4 for deep cultural fluency
- Uses a Korean-optimised tokenizer — up to ~2× faster/cheaper on Korean text
- Comes in THINK (reasoning), DASH (efficient) and open SEED variants
- Backed by Naver's gigawatt-scale sovereign-AI build-out with NVIDIA
Frequently asked questions
About this listing
This entry was compiled from publicly available data including HyperCLOVA X's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with HyperCLOVA X unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to HyperCLOVA X directly →
Spotted something out of date? Suggest an update →
Alternatives to HyperCLOVA X
More in LLMs & Chat
