CiQi-Agent-7B

CiQi-Agent

Github | Model | Dataset | Paper

---

CiQi-Agent: Aligning Vision, Tools and Aesthetics in Multimodal Agent for Cultural Reasoning on Chinese Porcelains

Accepted to ECCV 2026

---

🎯 Overview

CiQi-Agent has been accepted to ECCV 2026.

We present CiQi-Agent, a domain-specific multimodal agent for antique Chinese porcelain connoisseurship. The project is designed to combine fine-grained visual perception, tool-augmented reasoning, and cultural-heritage knowledge grounding for explainable porcelain analysis.

CiQi-Agent is built for tool-augmented multimodal reasoning on antique Chinese porcelains. During inference, it can inspect local visual evidence with an image zoom-in tool, retrieve visually similar examples with image retrieval, and access relevant domain knowledge with text retrieval, enabling more grounded and interpretable connoisseurship analysis.

Alongside the model, we release a single Hugging Face dataset repository that contains both the training data and benchmark data:

• CiQi-VQA, covering 29,596 porcelain specimens, 51,553 images, and 557,943 VQA pairs
• CiQi-Bench, the benchmark portion included in the CiQi-VQA repository, built from 775 porcelain specimens, 878 images, and 5,425 multiple-choice questions

On CiQi-Bench, CiQi-Agent achieves 81.5% average accuracy on multiple-choice evaluation and 66.7% average score on free-form evaluation, outperforming strong open-source and closed-source multimodal baselines.

🤖 Model

• Released as SII-Monument-Valley/CiQi-Agent-7B
• Built on Qwen2.5-VL-7B-Instruct with 8.29B parameters
• Architecture: Qwen2_5_VLForConditionalGeneration
• Inputs: text, one or more images, and optional tool definitions/results
• Outputs: text and optional Qwen-style tool calls
• Uses tool-augmented reasoning with zoom-in, image retrieval, and text retrieval
• Trained with a two-phase supervised fine-tuning + reinforcement learning pipeline
• Evaluated on CiQi-Bench with both multiple-choice and free-form protocols
• Achieves 81.5% average accuracy on multiple-choice evaluation, exceeding GPT-5 by 5.7 points and the strongest listed open-source baseline GLM-4.5V (72.6%) by 8.9 points
• Achieves 66.7% average score on free-form evaluation, exceeding GPT-5 by 18.7 points and Qwen2.5-VL-72B-Instruct by 23.7 points

The released checkpoint is stored in FP32 (approximately 33.2 GB) and can be downcast to BF16 by SGLang while loading. A conservative 4090 profile with a 4K context, an 8192-token KV cache pool, one running request, and CUDA graph disabled completed an OpenAI-compatible chat request using 17,676 MiB. The measurement was collected on a platform-provided 48 GB RTX 4090; it is below a 24 GB memory budget, but a standard 24 GB RTX 4090 was not directly tested. H200-class hardware remains recommended for longer contexts, higher concurrency, and production deployment.

🚀 Deployment

Install a recent SGLang release compatible with Qwen2.5-VL. The following command serves the model through an OpenAI-compatible API:

CUDA_VISIBLE_DEVICES=0 python -m sglang.launch_server \
  --model-path SII-Monument-Valley/CiQi-Agent-7B \
  --host 127.0.0.1 \
  --port 18901 \
  --served-model-name ciqi-agent \
  --tp 1 \
  --dp 1 \
  --dtype bfloat16 \
  --context-length 32768 \
  --mem-fraction-static 0.80 \
  --tool-call-parser qwen25 \
  --trust-remote-code \
  --enable-multimodal

For tensor parallel deployment, expose the required GPUs and set --tp to the same GPU count, for example CUDA_VISIBLE_DEVICES=0,1 with --tp 2.

Low-memory RTX 4090 profile

Use this configuration when targeting a 24 GB memory budget:

CUDA_VISIBLE_DEVICES=0 python -m sglang.launch_server \
  --model-path SII-Monument-Valley/CiQi-Agent-7B \
  --host 127.0.0.1 \
  --port 18901 \
  --served-model-name ciqi-agent \
  --tp 1 \
  --dp 1 \
  --dtype bfloat16 \
  --context-length 4096 \
  --max-total-tokens 8192 \
  --max-running-requests 1 \
  --chunked-prefill-size 1024 \
  --mem-fraction-static 0.88 \
  --disable-cuda-graph \
  --tool-call-parser qwen25 \
  --trust-remote-code \
  --enable-multimodal

This profile preserves BF16 model computation; --max-total-tokens limits KV-cache capacity rather than quantizing the model. A single conversation can use up to the configured 4096-token context, including text history, visual tokens, and output. Stop other GPU processes before startup and increase context, concurrency, or the KV-token pool only after checking available memory.

The low-memory configuration reduces context length and concurrency, not the precision of the loaded model. Further weight quantization may reduce memory use, but it can also affect multimodal reasoning and tool-call reliability and has not been validated for this release.

Do not expose the raw model server directly to the public internet. Put it behind TLS, authentication, rate limiting, request-size limits, logging, and key-revocation controls.

OpenAI-compatible chat

After the server becomes ready:

curl http://127.0.0.1:18901/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "ciqi-agent",
    "messages": [
      {"role": "user", "content": "Please introduce yourself in one sentence."}
    ],
    "temperature": 0,
    "max_tokens": 128
  }'

For an image reachable by the model server:

{
  "model": "ciqi-agent",
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "image_url",
          "image_url": {"url": "https://example.org/object.jpg"}
        },
        {"type": "text", "text": "Describe the object and explain the visual evidence."}
      ]
    }
  ]
}

Tool calling

The model follows the Qwen2.5 tool-calling convention. When serving with SGLang, keep --tool-call-parser qwen25 enabled and pass tools through the OpenAI-compatible tools field. Tool execution remains the responsibility of the client or agent runtime; return each tool result to the model before asking it to continue.

📊 Model Performance

Multiple-Choice Accuracy (%) on CiQi-Bench

Free-Form Score (%) on CiQi-Bench

📦 Dataset & Benchmark

The SII-Monument-Valley/CiQi-VQA repository contains both the CiQi-VQA training data and the CiQi-Bench evaluation data.

📊 CiQi-VQA Training Data

CiQi-VQA is a large-scale dataset for porcelain-centered multimodal training.

• 29,596 porcelain specimens
• 51,553 images
• 557,943 VQA pairs
• 38 dynasties
• 42 reign periods
• 246 glaze color categories
• 248 decorative motif categories
• 158 vessel shape categories

Link: https://huggingface.co/datasets/SII-Monument-Valley/CiQi-VQA

🧪 CiQi-Bench Evaluation Data

CiQi-Bench is the expert-aligned benchmark portion of the same CiQi-VQA repository for evaluating porcelain connoisseurship ability.

• 775 porcelain specimens
• 878 images
• 5,425 multiple-choice questions
• Free-form evaluation with attribute-wise scoring

Repository: https://huggingface.co/datasets/SII-Monument-Valley/CiQi-VQA

📈 Dataset and Benchmark Statistics

Intended Use and Limitations

The model is intended for research and non-commercial cultural-heritage visual question answering, multimodal reasoning, retrieval-augmented generation, and agent/tool-use experiments.

• Outputs may be factually incorrect and should not be treated as provenance, authentication, attribution, appraisal, conservation, or legal advice.
• The model may inherit biases and limitations from its base model and training data.
• Users should verify museum, collection, licensing, and attribution metadata against authoritative sources.
• Static repository scans found no plaintext credentials or private deployment paths in this release. This does not prove that a neural model cannot reproduce fragments of its training data; deployments handling sensitive inputs should apply independent memorization and extraction-risk testing.
• Do not send confidential images or text to a third-party deployment unless its retention and access policies are acceptable.

📜 License

• Model license: CC BY-NC 4.0
• Dataset license: CC BY-NC 4.0

Users are also responsible for complying with the licenses and terms of the base model, source images, and any external retrieval corpus they use.

🤝 Acknowledgement

We thank Verl for providing an open-source reinforcement learning framework that supports this line of research.

We also thank DeepEyes for inspiring and informing our exploration of tool-augmented multimodal reasoning.

📜 Citation

@inproceedings{wang2026ciqiagent,
      title={CiQi-Agent: Aligning Vision, Tools and Aesthetics in Multimodal Agent for Cultural Reasoning on Chinese Porcelains},
      author={Wenhan Wang and Zhixiang Zhou and Zhongtian Ma and Yanzhu Chen and Ziyu Lin and Hao Sheng and Pengfei Liu and Honglin Ma and Wenqi Shao and Qiaosheng Zhang and Yu Qiao},
      booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
      year={2026},
      eprint={2603.28474},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.28474},
}