Atlas

Models

← All models

HyperCLOVA X SEED 32B Think

NAVER · 2025-12-29 · 33.3B parameters

HyperCLOVA X SEED 32B Think is NAVER Cloud's open-weight vision-language reasoning model released on December 29, 2025; the complete checkpoint contains exactly 33,313,410,304 parameters, including its Qwen2.5-VL-derived vision encoder and linear projector. Its 72-layer dense decoder accepts text, images, and video in a 131,072-token configured context, and its chat template exposes optional thinking and non-thinking generation while training emphasizes Korean linguistic and cultural context, multimodal understanding, and agentic behavior.

Benchmark scores

BenchmarkScore
AA-LCR11.7
AA-Omniscience Index-52.9
AIME 202559.0
Artificial Analysis Intelligence Index17.0
Artificial Analysis Omniscience Accuracy14.6
Artificial Analysis Omniscience Hallucination Rate79.1
Artificial Analysis Openness Index30.6
Capability134.1
CritPt0.0
GPQA Diamond61.5
Humanity's Last Exam (Text-Only)5.5
IFBench (Artificial Analysis)37.9
LiveCodeBench (Artificial Analysis)62.9
MMLU Pro78.5
SciCode (Artificial Analysis)28.4
Terminal-Bench Hard12.1
τ²-bench (Telecom)87.4
Loading Atlas data…