HyperCLOVA X

Naver splits the family by job rather than by version number. THINK is the reasoning tier — multimodal, staged reasoning across text and images, aimed at STEM problems and instruction-following against visual input. DASH is the lightweight, low-latency tier for high-volume work where cost per call dominates. SEED is the open-source tier, and it is the one most readers of this page can actually use.

The SEED line as published on Hugging Face:

• SEED Think 32B — a 33B-parameter vision-language model with a unified backbone that puts text tokens and visual patches in a shared embedding space. 128K context, text/image/video in, text out, optional thinking mode with a configurable token budget, knowledge cutoff May 2025.
• SEED Think 14B — the earlier reasoning model in the same line.
• SEED Omni 8B — the family's any-to-any model, handling text, audio and vision as both inputs and outputs in a single set of weights rather than a chain of modality-specific pipelines. Naver published a technical paper on it in January 2026.
• SEED Vision Instruct 3B — image and video understanding tuned for Korean-language and Korean-cultural content.
• SEED Text Instruct 1.5B and 0.5B — small text models for translation between Korean and English or Japanese, and for multi-turn Korean conversation in resource-constrained places like mobile apps, wearables and smart home devices.

Hardware expectations are documented rather than implied. Naver's model card for SEED Think 32B asks for roughly 68GB of VRAM across three GPUs, or a 4×A100 80GB configuration, and ships OmniServe — its own production inference server with an OpenAI-compatible API — alongside the weights. Quantised community builds for llama.cpp, Ollama and LM Studio exist.

These are open weights, not open-source in the Apache sense. Every SEED model carries a HyperCLOVA X SEED Model License Agreement written by Naver. The organisation card states plainly that the models are available for commercial use by businesses, researchers and developers, and fine-tuning is expected — there are already community fine-tunes and quantisations on the Hub. But a bespoke licence means bespoke terms, and a legal review that clears Apache 2.0 does not automatically clear this. Read the LICENSE file in the specific repository.

The hosted route is CLOVA Studio, Naver Cloud's no-code and API development environment for the family. Naver says more than 2,000 companies and institutions use it. Beyond raw completions it bundles a playground, vision models that read documents and tables, function calling, a Router that performs precise classification from a natural-language prompt with no tuning required, SFT tuning against your own dataset, a Skill Trainer for wiring the model to Naver or third-party APIs, and Explorer tools covering embeddings, summarisation, reranking, paragraph splitting and RAG reasoning.

Two heavier tiers sit above it. CLOVA Studio Exclusive pairs SFT tuning with dedicated inference infrastructure, sold on the promise of guaranteed throughput. Neurocloud for HyperCLOVA X is the unusual one: a hybrid-cloud hardware appliance installed inside the customer's own data centre, combining a GPU cluster with HyperCLOVA X training and operating tools, built for organisations whose security rules simply forbid sending prompts outside the building. That is a product Korean banks, hospitals and government agencies buy, and it explains a lot of the deployment list — the Seoul Metropolitan Government's public data chatbot, the National Information Society Agency's AI-Hub assistant, a legal Q&A service for the DR & AJU law group, a health-checkup explainer for KMI, and Naver Works' own WORKS AI.

CLOVA Studio bills per 1,000 tokens in Korean won, excluding VAT, with the rate depending on which engine you call. The newer models — HCX-007, HCX-005 and HCX-DASH-002 — separate input and output pricing; the older HCX-003 and HCX-DASH-001 do not. Tuning runs are billed separately from inference, as are skill sets, the router and each Explorer tool. Naver's public product page lays out that structure in full but shows the actual figures as blanks, so the number you will pay comes from the pricing calculator or the console after sign-in rather than from an open rate card. Anyone comparing this against a published $/M-token price list should budget time for a sales conversation.

The one cost lever Naver does argue in public is tokenisation. The family uses a Korean-optimised tokenizer, and Naver claims it therefore runs up to twice as fast and at lower cost than English-centric models on Korean text. Fewer tokens per Korean sentence is a real structural advantage, and it compounds on long-context workloads.

Naver's own comparisons — 84.14 versus 68.2 on HAE-RAE Bench, 70.7 versus 38.6 on KorNAT common-knowledge alignment, 93.8 versus 89.0 for GPT-4 on English-to-Korean translation — are vendor-published and date from the original HyperCLOVA X launch materials, so treat them as a claim rather than a finding. The stronger current evidence is third-party: Artificial Analysis scores SEED Think 32B at 44 on its Intelligence Index and calls it one of the strongest South Korean models, ahead of LG's EXAONE 4.0 32B, with 82% on the Korean split of Global MMLU Lite.

That is the right frame. HyperCLOVA X competes in the Korean field, not the global one, and the Korean field is getting crowded — Naver notably sat out the government's free "AI for All" programme in August 2026 while the telecom operators and Kakao bid for it. If you need a Korean-language model you can download, license commercially, fine-tune and run on-premises, this family is a serious candidate and the SEED weights cost nothing to try. If you need the best model available and Korean is just one of the languages involved, the frontier labs remain the answer.