HyperCLOVA X THINK

Naver’s HyperCLOVA X THINK is the first reasoning-centric model in the HyperCLOVA X family, built for deep step-by-step problem solving in Korean and English. The technical report describes pretraining on roughly 6 trillion high-quality tokens and a compute-/memory-balanced Peri-LN Transformer scaled with μP. A three-stage curriculum expands the context window to 128K tokens. Post-training combines supervised fine-tuning with reinforcement learning from verifiable rewards, and the model can operate in either “detailed rationale” or concise-answer modes.

Beyond text, HyperCLOVA X THINK integrates visual inputs for multimodal reasoning—e.g., STEM problems and image-based instruction following—while emphasizing strong Korean language competence for agentic applications. Naver’s announcement materials highlight top performance on domestic expert-level Korean evaluations and note intentions to open-source components of the reasoning model.