Instructions to use KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B", trust_remote_code=True)
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B

SGLang

How to use KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B with Docker Model Runner:
```
docker model run hf.co/KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B
```

HyperCLOVA X KETI-HAECHI-32B

HAECHI personalization with Qwen and HyperCLOVA

Powered by HyperCLOVA X

HyperCLOVA X KETI-HAECHI-32B is a multimodal model derived from naver-hyperclovax/HyperCLOVAX-SEED-Think-32B. It was developed for two primary purposes: improving Korean cultural-heritage understanding and Korean OCR, and improving tool calling and multi-step, stateful agent execution. Alongside these goals, the model retains broad multimodal, language, and coding capabilities from the base model.

Core capability profile

Korean cultural heritage and OCR: identifies the official names of heritage objects, answers questions grounded in heritage images, and reads Korean text from signs, scenes, rendered text, and public documents.
Tool calling and long-horizon task execution: selects and calls tools, carries information across multiple turns, tracks changing state, and works toward an end-to-end goal over several steps.
General multimodal understanding: interprets images and text together, follows Korean and English instructions, and performs visual reasoning beyond the specialized heritage domain.

Performance overview

Bar chart comparing the official HyperCLOVA X SEED 32B Think base model with HyperCLOVA X KETI-HAECHI-32B on representative cultural-heritage, Korean OCR, long-horizon agent, tool-calling, and general multimodal benchmarks

The chart highlights representative metrics from each major capability area. Detailed benchmark tables appear in the Evaluation section.

Quickstart

The checkpoint contains custom HyperCLOVA X model and processor code, so trust_remote_code=True is required. The example below follows the direct Transformers runtime used to verify the packaged BF16 checkpoint on two 80 GB GPUs. torchvision is required when AutoProcessor initializes the bundled video processor, including for image-only inference. The versions below are the combination verified in a clean environment.

pip install "torch==2.9.1" "torchvision==0.24.1" "transformers>=4.57,<5" accelerate pillow safetensors huggingface_hub

import sys

import torch
from huggingface_hub import hf_hub_download
from PIL import Image
from transformers import AutoProcessor
from transformers.dynamic_module_utils import get_class_from_dynamic_module

model_id = "KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B"
image_path = hf_hub_download(
    repo_id=model_id,
    filename="samples/qualitative/01_gyeongbokgung_gyeonghoeru.jpg",
)

processor = AutoProcessor.from_pretrained(
    model_id,
    trust_remote_code=True,
    use_fast=False,
)

# The bundled visual code selects FlashAttention2 on Ampere-class GPUs.
# This verified fallback uses Transformers SDPA when flash-attn is unavailable.
model_class = get_class_from_dynamic_module(
    "modeling_vlm.HCXVisionForCausalLM",
    model_id,
)
sys.modules[model_class.__module__].is_ampere_or_newer = lambda: False

model = model_class.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="balanced",
    max_memory={0: "76GiB", 1: "76GiB"},
    low_cpu_mem_usage=True,
).eval()

image = Image.open(image_path).convert("RGB")
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {"url": image_path},
            },
            {
                "type": "text",
                "text": "사진 속 국가유산의 정확한 공식 명칭만 답하세요. 설명은 쓰지 마세요.",
            },
        ],
    }
]

prompt = processor.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    thinking=False,
)
inputs = processor(
    text=[prompt],
    images=[image],
    min_pixels=65_536,
    max_pixels=1_048_576,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output_ids = model.inference(
        **inputs,
        max_length=32,
        min_length=0,
        temperature=0.0,
        do_sample=False,
        use_cache=True,
    )

answer = processor.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
print(answer)

# Example output from this checkpoint:
# 경복궁 경회루

For the packaged directory in this release, use model_id = "./model" and open the sample image directly from samples/qualitative/01_gyeongbokgung_gyeonghoeru.jpg. Adjust device_map and max_memory for the deployment hardware. The upstream model also supports the OmniServe OpenAI-compatible runtime.

The example disables thinking mode for concise extraction. Set thinking=True and provide a larger generation budget for tasks that benefit from the model's reasoning mode.

Evaluation

All reported deltas are calculated as HyperCLOVA X KETI-HAECHI-32B minus the untouched official HyperCLOVA X SEED 32B Think weights. Percentage deltas are absolute percentage points. The evaluator-compatible base copied only the chat template; model weights were unchanged. Both models used matched API, template, decoding, stop-token, parser, and dataset settings.

Qualitative comparison with the base model

The examples below were re-run from the packaged BF16 checkpoint and the untouched official base weights using HCXVisionForCausalLM.inference() with the original evaluation prompts. Both models used thinking disabled, greedy decoding, a 256-token output limit, and the same image preprocessing settings. For OCR examples, the table shows the recognized text span; surrounding explanatory prose is omitted.

Korean cultural-heritage recognition

Target (reference)	HyperCLOVA X SEED 32B Think base	HyperCLOVA X KETI-HAECHI-32B
금동연가7년명여래입상	❌ `금동미륵보살반가사유상 (국보 제78호)`	✅ `금동연가7년명여래입상`
백자 철화포도원숭이문 항아리	❌ `백자철화포도문항아리`	✅ `백자 철화포도원숭이문 항아리`
무령왕 금제 관식	❌ `금제 관식`	✅ `무령왕 금제 관식`

Korean OCR

Font

Target (reference)	HyperCLOVA X SEED 32B Think base	HyperCLOVA X KETI-HAECHI-32B
`왠`	❌ `왜`	✅ `왠`
`저절로`	❌ `절로`	✅ `저절로`
`입다`	❌ `입니다`	✅ `입다`
`찹`	❌ `참`	✅ `찹`

Outdoor

Target (reference)	HyperCLOVA X SEED 32B Think base	HyperCLOVA X KETI-HAECHI-32B
`동전 노래연습장`	❌ `동전 노래방`	✅ `동전 노래연습장`
`뚱스`	❌ `뚠스`	✅ `뚱스`

Public exec

Target (reference)	HyperCLOVA X SEED 32B Think base	HyperCLOVA X KETI-HAECHI-32B
`재난안전관리과`	❌ `폐쇄병동전과`	✅ `재난안전관리과`
`김해시장(전산정보과장)`	❌ `개인 정보에 해당합니다.`	✅ `김해시장(전산정보과장)`

Quantitative benchmarks

Korean cultural-heritage performance

Benchmark	Official base	HyperCLOVA X KETI-HAECHI-32B	Delta
H400 direct exact	5.05%	74.16%	+69.11 pp
H400 hard choice	59.79%	89.76%	+29.97 pp
H400 knowledge image	4.89%	35.13%	+30.24 pp
H400 knowledge text	9.07%	39.59%	+30.53 pp

H400 is an internal evaluation set built around 400 Korean cultural-heritage items. It evaluates canonical-name recognition, fine-grained discrimination among visually similar heritage items, and heritage knowledge grounded in either images or text. H400 direct exact measures open-ended canonical-name retrieval, while H400 hard measures closed-set discrimination. The knowledge-image and knowledge-text values are mean string-similarity scores in this report.

Korean (Hangul) OCR performance

Benchmark	Official base	HyperCLOVA X KETI-HAECHI-32B	Delta
`ocr_font` exact	18.36%	20.96%	+2.60 pp
`ocr_outdoor` exact	33.59%	38.87%	+5.27 pp
`ocr_public_exec` exact	2.47%	1.69%	-0.78 pp

These Mammoth OCR results use strict exact match. Font and outdoor OCR improve, while public-document exact match declines slightly despite selected successful examples above.

Tool calling and long-horizon task performance

Tau2

Benchmark	Official base	HyperCLOVA X KETI-HAECHI-32B	Delta
Tau2 airline (n=32)	46.88%	34.38%	-12.50 pp
Tau2 retail (n=32)	37.50%	53.12%	+15.62 pp
Tau2 telecom (n=32)	37.50%	40.62%	+3.12 pp
Tau2 weighted overall (n=96)	40.62%	42.71%	+2.08 pp

Tau2 measures end-to-end success across multi-step airline, retail, and telecom workflows. The weighted overall score improves through retail and telecom gains, although airline performance declines.

BFCL V4

Benchmark	Official base	HyperCLOVA X KETI-HAECHI-32B	Delta
Overall accuracy	7.85%	15.00%	+7.15 pp
Non-Live AST accuracy	21.42%	64.17%	+42.75 pp
Non-Live Simple AST	23.17%	48.67%	+25.50 pp
Non-Live Multiple AST	45.50%	90.50%	+45.00 pp
Non-Live Parallel AST	13.50%	60.50%	+47.00 pp
Non-Live Parallel Multiple AST	3.50%	57.00%	+53.50 pp
Multi-turn accuracy	2.63%	13.63%	+11.00 pp
Multi-turn base	4.00%	19.50%	+15.50 pp
Multi-turn missing function	2.50%	9.50%	+7.00 pp
Multi-turn missing parameter	1.50%	10.50%	+9.00 pp
Multi-turn long context	2.50%	15.00%	+12.50 pp

BFCL V4 measures structured function-call generation and multi-turn recovery from missing functions or parameters. The tuned checkpoint improves across all reported BFCL V4 categories.

General multimodal and language retention

Benchmark	Official base	HyperCLOVA X KETI-HAECHI-32B	Delta
General-VL macro (7)	66.19%	65.29%	-0.90 pp
MMBench DEV EN v1.1	83.36%	83.59%	+0.23 pp
MMStar	63.00%	63.13%	+0.13 pp
MMStar-KO	59.27%	60.00%	+0.73 pp
KRETA	84.32%	78.08%	-6.25 pp
MMMU-Pro 10c	38.15%	38.96%	+0.81 pp
HallusionBench aAcc	66.46%	64.77%	-1.68 pp
MathVista MINI	68.80%	68.50%	-0.30 pp
OpenCompass Core	54.98%	54.20%	-0.78 pp
OpenCompass Extra	68.67%	68.07%	-0.60 pp
OpenCompass Korean	53.74%	53.01%	-0.73 pp

See the detailed score report for the full results, the benchmark guide for definitions and caveats, and the paired output report for additional raw Base/Tuned examples.

Intended use

The model is intended for research and prototyping involving:

Korean cultural-heritage image identification and visual question answering;
Korean scene, sign, font, and public-document OCR;
image-grounded heritage knowledge retrieval;
multi-step, stateful tool-use workflows with explicit monitoring and recovery;
general image understanding and text generation; and
analysis of domain specialization and capability retention in multimodal models.

Responsible use

Verify cultural-property names and factual claims against authoritative catalogs or domain experts. Preserve human review for public descriptions, education, archival metadata, and research outputs. Do not treat model output as evidence of authenticity or provenance. Avoid submitting sensitive or personal documents for OCR unless the deployment provides appropriate privacy controls. All use must comply with the Acceptable Use Policy incorporated into the HyperCLOVA X model license.

Limitations

OCR outputs may omit text, normalize spelling incorrectly, or add unsupported text. Public-document exact match is lower than the official base in this evaluation.
The model can hallucinate names, dates, designations, provenance, and historical claims. Similar-looking artifacts and uncommon viewpoints are especially challenging.
General-VL macro, KRETA, HallusionBench, and the Tau2 airline domain regress relative to the official base.
Direct Transformers loading relies on bundled custom code and is sensitive to the Transformers version, attention backend, and available GPU memory.
General multilingual, video, robustness, demographic-bias, privacy, and safety behavior were not comprehensively evaluated for this release.
As with the base model, generated content may be inaccurate, biased, unsafe, or unsuitable for the user's context.

Core contributors

San Kim (kimsan0622@keti.re.kr) — Led long-context agent task performance and development/management of the main training loop.
Byunggill Joe (byunggill@keti.re.kr) — Secured GPU compute resources through the support program and led Korean cultural-heritage recognition/OCR performance.

Acknowledgements

HyperCLOVA X KETI-HAECHI-32B was developed using compute resources provided through the 첨단 GPU 활용 지원 사업 of the National IT Industry Promotion Agency (NIPA; 정보통신산업진흥원).

License and attribution

This derivative checkpoint is distributed under the HyperCLOVA X SEED 32B Think Model License Agreement, not the Apache License 2.0. Redistribution and use are subject to that agreement and its Acceptable Use Policy, including naming, attribution, notice, and display requirements. See NOTICE for the required attribution and modification notice. Third-party data and sample images may be subject to additional terms.

Powered by HyperCLOVA X

Citation

@misc{hyperclova_x_keti_haechi_32b,
  title        = {HyperCLOVA X KETI-HAECHI-32B},
  author       = {Kim, San and Joe, Byunggill},
  year         = {2026},
  howpublished = {Hugging Face model repository},
  url          = {https://huggingface.co/KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B}
}

한국어 요약

이 모델은 naver-hyperclovax/HyperCLOVAX-SEED-Think-32B를 기반으로 두 가지 목적을 위해 개발한 멀티모달 모델입니다. 첫 번째 목적은 한국 문화유산의 정확한 명칭 식별, 이미지 기반 문화유산 질의응답, 한글 OCR 능력을 향상하는 것입니다. 두 번째 목적은 적절한 도구를 호출하고 여러 단계에 걸쳐 정보를 기억하며 변화하는 상태를 추적하는 장기작업 수행 능력을 향상하는 것입니다.

공식 베이스 모델과 동일한 조건의 내부 평가에서 H400 직접 식별 정확도는 5.05%에서 74.16%로, H400 유사 문화재 선택 정확도는 59.79%에서 89.76%로 향상되었습니다. 한글 OCR은 글꼴과 실외 간판에서 향상되었지만 공공문서 exact match는 2.47%에서 1.69%로 하락했습니다. 장기작업 평가인 Tau2 가중 점수는 40.62%에서 42.71%로, BFCL V4 전체 정확도는 7.85%에서 15.00%로 향상되었습니다. 실제 배포에서는 단계별 상태 확인, 실패 복구, 결과 검증을 함께 적용하는 것을 권장합니다.

Downloads last month: 30

Safetensors

Model size

33B params

Tensor type

BF16

Model tree for KETI-AIR/HyperCLOVA-X-KETI-HAECHI-32B

Base model

naver-hyperclovax/HyperCLOVAX-SEED-Think-32B

Finetuned

(4)

this model