HyperCLOVA X
NAVER's Korean-first LLM family, sold to developers as the CLOVA Studio API on NAVER Cloud Platform. The line spans HCX-007 for hybrid reasoning, HCX-005 for multimodal image understanding and HCX-DASH-002 as the lightweight throughput tier.
Pros
- The strongest published Korean-language performance of any commercial LLM, trained on roughly 6,500 times more Korean data than GPT-4 and leading K-MMLU, HAE-RAE Bench and KorNAT
- a tokenizer built for Korean cuts token counts on Korean text roughly in half, so real Korean workloads cost meaningfully less than the headline rate implies. HCX-007 and HCX-005 both carry 128K context with function calling and structured outputs, tuning runs from PEFT up to SFT and RLHF on the Exclusive plan, and Neurocloud allows a full hybrid deployment inside a customer's own data centre.
Cons
- Per-token rates are not published on any public page: they sit behind the NAVER Cloud Platform portal under Services then AI Services then CLOVA Studio, so you cannot price a project before opening an account, and both the Exclusive and Neurocloud tiers are contact-sales only. English documentation exists and is decent, but the console, billing and support remain Korea-centric, and NAVER makes no claim to benchmark leadership outside Korean, Japanese and Chinese. The older HCX-003 and HCX-DASH-001 still on the list top out at 8,192 and 4,096 tokens of context.
Cost
No public per-token price list. The Basic plan is pay-as-you-go on token usage with separate input and output rates for HCX-007, HCX-005 and HCX-DASH-002, published only inside the NAVER Cloud Platform portal after signup. Exclusive is a subscription priced on tokens per minute with dedicated infrastructure, and Neurocloud for HyperCLOVA X is a hybrid deployment managed in the customer's own data centre; both are quoted by sales. NAVER states the current generation improved operating cost by more than 50% over its predecessor, but publishes no figure to check that against.
Verdict
The right call when Korean-language quality, Korean data residency or integration with NAVER's own ecosystem is the actual requirement; for anything else, the opaque pricing and Korea-centric onboarding are hard to justify against a globally available API you can price before you sign up.
Visit HyperCLOVA X