Model
Model explorer

NAVER HyperCLOVA X SEED Think 32B

OPEN
NAVER · HyperCLOVA X SEED family · released Dec 26, 2025

Open-weights multimodal reasoning model; strong Korean open model, reported ahead of EXAONE 4.0 32B on Artificial Analysis Intelligence Index at launch (44 vs 43; later shown as 17 under recalibrated v4.1 methodology).

ReasoningCodingVisionFunction callingTool useAgentic
894.7
Elo · rank #235
Parameters
32B
Active params
32B (dense)
Context
128K tokens
Architecture
Dense multimodal (vision-language) transformer with optional extended-thinking mode
License
HyperCLOVA X SEED 32B Think Model License Agreement
Languages
—
API price (in/out)
No hosted API
Modalities
text · vision · video
Benchmark results
Bar shows position within the tracked field; marker = field best
DocVQAVision95.5%#5
best: Qwen2-VL-72B · 96.5%
GPQA DiamondReasoning62.0%#157
best: GPT-5.6 · 94.6%
HellaSwagReasoning57.2%#144
best: Claude 3 Opus · 95.4%
Humanity's Last ExamReasoning5.0%#119
best: Claude Sonnet 5 · 57.4%
IFBenchReasoning38.0%#55
best: MiniMax M3 · 83.0%
MMLUKnowledge87.7%#28
best: OpenAI o3 · 92.9%
PIQAReasoning76.7%#65
best: GPT-4o mini · 93.1%
SEED-BenchVision77.9%#1
best: this model · 77.9%
τ²-Bench AirlineAgents58.0%#5
best: Trinity-Large-Thinking · 88.0%
τ²-Bench RetailAgents71.6%#9
best: Claude Sonnet 4.6 · 91.7%
τ²-Bench TelecomAgents65.1%#30
best: Claude Opus 4.6 · 99.3%
Terminal-BenchCoding21.8%#15
best: Claude Opus 4.5 (High) · 59.3%
TextVQAVision85.4%#3
best: Molmo 2 8B · 85.7%
Video-MMEVision63.4%#20
best: Seed 2.1 Pro · 89.2%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
66 GB
Fits on (Q4)
Multi-node cluster required
No community GGUF/MLX quantization confirmed yet as of research date; custom VLM architecture not yet supported by llama.cpp
Quantizations
—
HyperCLOVA X SEED family
Elo progression across releases
NAVER HyperCLOVA X SEED Think 32BDec 2025894.7
Resources
HFnaver-hyperclovax/HyperCLOVAX-SEED-Think-32BDOCSartificialanalysis.ai/models/hyperclova-x-seed-think-32b
API price weights · each benchmark row carries its own source badge (see methodology)