HyperCLOVA X SEED Think

Vision-language reasoning model with chain-of-thought and agentic capabilities. 32B parameters, 128K context. Supports text, image, and video input. AA Intelligence Index v4.1: 17 (AA-estimated). Open weights with commercial license. Also available as 14B variant (text-only, pruned + distilled from Naver's production HCX-007).

Provenance: the text backbone is Naver's own from-scratch lineage — a custom 72-layer HyperCLOVAX architecture with Naver's in-house tokenizer, descended from the ~6T-token HyperCLOVA X THINK pretrain; no Qwen shape or tokenizer match. The 32B's vision encoder, however, adopts Qwen2.5-VL's ViT — Jan 2026 Korean press reported >99% weight similarity, fueling a sovereign-AI controversy. Naver maintains the reasoning core is fully proprietary.