Naver's HyperCLOVA X now reaches users three ways: six free SEED repos on Hugging Face, managed CLOVA Studio on NAVER Cloud Platform, and features inside Naver's own services, since the CLOVA X chatbot closed on April 9, 2026. Its license adds conditions above 10 million monthly users.
What changed
HyperCLOVA X is Naver's Korean-first foundation model family. Its flagship open-weight SEED 32B Think model, released December 29, 2025, reads text, images and video within a 128,000-token context and can be self-hosted for free, including commercially, under Naver's own license.
Provider: Naver · Family: HyperCLOVA X
Context window: 128,000 tokens
Input modalities: text, image, video · Output: text
About HyperCLOVA X
HyperCLOVA X is the foundation model family built by Naver, South Korea's largest internet company, and its cloud arm Naver Cloud. Naver launched the original HyperCLOVA X model in August 2023, an 82-billion-parameter dense model reached through Naver Cloud's CLOVA Studio platform. Since then the family has split into three lines: HyperCLOVA X THINK, a reasoning-focused line; HyperCLOVA X SEED, an open-weight line Naver releases for research and commercial use as part of its sovereign AI strategy; and HyperCLOVA X DASH, a lightweight model built for fast, high-throughput tasks. HyperCLOVA X sits at the center of Naver's push to keep Korean-language AI infrastructure, data and models inside Korea rather than depending on foreign providers.
The newest open release is HyperCLOVA X SEED 32B Think, which Naver Cloud published on December 29, 2025 together with HyperCLOVA X SEED 8B Omni, as the first output of its project under the Ministry of Science and ICT's Independent AI Foundation Model initiative. The Hugging Face model card describes SEED 32B Think as a dense vision-language model that takes text, image and video input, holds up to 128K tokens of context, offers an optional thinking mode and has a May 2025 knowledge cutoff; Hugging Face counts 33.3 billion parameters once the vision encoder is included. Naver tests its agent skills on the Tau^2 airline, retail and telecom tool-use suites. The 8B Omni model learns text, images and audio inside one network and can generate and edit images from a text prompt, according to Naver's launch announcement.
An earlier reasoning model is documented in Naver's June 2025 HyperCLOVA X THINK technical report (arXiv:2506.22403), which describes pretraining on roughly 6 trillion high-quality Korean and English tokens. Naver reports the vision-augmented THINK model scoring at or above GPT-4.1 on Korea's CSAT STEM exam, and cites competitive results on KMMLU, CSAT, KoBALT-700, HAERAE-1.0 and KoBigBench; these are vendor-reported figures on Korean-language test sets, not independent results on SWE-bench, GPQA or MMLU-Pro.
The open SEED line spans six Hugging Face repositories under Naver's naver-hyperclovax organization: two text-only models at 0.5B and 1.5B parameters, a 3B vision-instruct model, the 8B omni-modal model, and the 14B and 32B Think reasoning models. They ship under Naver's own HyperCLOVA X SEED license rather than an OSI licence. It permits free commercial use with conditions: a product built on the model must display "Powered by HyperCLOVA X", a model trained on its outputs must start its name with "HyperCLOVA X", and a company whose services passed 10 million monthly active users, or that offers a product competing directly with Naver's, must request a separate license from Naver (Section 4 of the SEED 32B Think license).
There are three ways to use HyperCLOVA X today. Developers download SEED weights from Hugging Face and serve them with OmniServe, Naver's open-source inference server, which exposes an OpenAI-compatible API. Enterprises call the managed models through CLOVA Studio on NAVER Cloud Platform, or run them on Neurocloud for HyperCLOVA X, a hybrid setup on dedicated hardware. Consumers no longer have a standalone chat app: Naver closed the CLOVA X chatbot and its Cue: AI search on April 9, 2026, and now puts its AI work into agents inside its own search and shopping services. CLOVA Studio has no public per-token rate card as of this review; billing is set in the NAVER Cloud Platform console.
HyperCLOVA X's clearest advantage is depth of Korean-language optimization, backed by direct distribution through Naver's own search, commerce and content products. Its clearest gap is external validation: Naver's December 2025 announcement cites an Artificial Analysis composite index, but the company has not published its own SWE-bench Verified, GPQA Diamond, AIME, MMLU-Pro or ARC-AGI-2 results, so buyers comparing HyperCLOVA X with GPT, Claude or Gemini on general-purpose reasoning have to run their own tests.
Inside Korea the closest rivals are Upstage's Solar Pro 4, LG AI Research's K-EXAONE 2.0 and Kakao's Kanana-2 30B-A3B Instruct, all aimed at Korean-language quality. A team that needs a capable open model at this size but not Korean-specific tuning may be better served by Gemma 4 31B or Qwen3.6-27B: both ship under Apache 2.0, with no user threshold and no attribution banner, and both have wider third-party benchmark coverage.
Pricing
The SEED line can be downloaded and run on your own infrastructure at no cost. A managed CLOVA Studio option exists for teams that would rather not self-host; check the NCP console for current billing figures, since no public rate card covers it.
Key Features
- HyperCLOVA X SEED Think: Naver's flagship open-weight reasoning model under the HyperCLOVA X SEED license; Hugging Face lists 33.3 billion parameters including its vision encoder.
- Text, image and video input: The Think model reads text, images and video in one shared embedding space, and Naver tests its tool use on 3 agent suites: airline, retail and telecom.
- Six-model SEED family: 6 open Hugging Face repos cover text-only, vision-instruct, omni-modal and reasoning configurations, so a team can match model size to its GPU budget.
- Controllable thinking mode: A thinking_token_budget setting caps how many reasoning tokens the Think model spends, with a default of 500 in Naver's OmniServe server.
- CLOVA Studio managed access: Enterprises that prefer not to self-host can reach HyperCLOVA X through Naver Cloud's CLOVA Studio platform, which Naver says more than 1,000 businesses and institutions already use.
Pros
- Free, open-weight self-hosting across the entire SEED line, from the smallest text model up to the flagship Think reasoning model, including commercial use under Naver's license.
- Korean-first pretraining, with vendor-reported results on Korean exam and knowledge benchmarks that few global labs test against.
- One open model handles text, images and video within one long context window, so long Korean documents and screenshots need no separate vision model.
Cons
- Naver publishes no SWE-bench, GPQA or MMLU-Pro results, and this review found no independent head-to-head test against GPT, Claude or Gemini.
- CLOVA Studio's managed API does not publish flat per-token USD or KRW pricing, unlike most competing model APIs.
- Not an OSI licence: products must show a "Powered by HyperCLOVA X" notice, and large or Naver-competing companies need Naver's separate permission.
Frequently Asked Questions
How much does HyperCLOVA X cost to use?
The open HyperCLOVA X SEED weights, from the smallest text model to the flagship Think reasoning model, cost nothing to download and self-host, commercial use included. Naver Cloud also sells managed access through CLOVA Studio, billed via the NAVER Cloud Platform console; NCP has not published a flat per-token USD or KRW rate.
How does HyperCLOVA X compare on benchmarks vs GPT-4.1?
Naver's HyperCLOVA X THINK technical report says its vision-augmented THINK model matches or exceeds GPT-4.1 on the Korean CSAT STEM benchmark, with competitive scores on Korean tests such as KMMLU and HAERAE-1.0. These are vendor-reported, Korean-language comparisons; nobody has published independent Tier-1 global results against GPT-4.1.
Is HyperCLOVA X open source or proprietary?
The SEED line is open-weight but not OSI open source: six models sit on Hugging Face under Naver's own license, which allows commercial use with a "Powered by HyperCLOVA X" notice. The original 2023 flagship and the THINK and DASH managed variants remain proprietary, reached through CLOVA Studio.
Does HyperCLOVA X train on user data?
Naver has not published a specific policy on whether CLOVA Studio's managed HyperCLOVA X access trains on customer inputs. Data handling for managed access is set out in CLOVA Studio's terms of service on the NAVER Cloud Platform console; self-hosted SEED deployments run entirely on infrastructure the user controls.
Who is HyperCLOVA X best for and who should avoid it?
Teams building Korean-language products, or organizations needing in-country AI deployment in South Korea, get Korean-first tuning and free self-hosting. Companies past the license's monthly-user threshold, anyone needing an Apache-style license, and buyers who want a published global Tier-1 benchmark comparison should look at an alternative.