We provide our fine-tuned model weights openly via ModelScope. AirIAD is built upon the powerful Qwen3-VL-Instruct architecture, offering two scalable versions to balance deployment efficiency and reasoning capability.
| Model Version | Base Model | Parameters | ModelScope Link |
|---|---|---|---|
| AirIAD-4B | Qwen3-VL-Instruct-4B | 4B | 🔗 FisherYuuri/AirIAD-4B |
| AirIAD-8B | Qwen3-VL-Instruct-8B | 8B | 🔗 FisherYuuri/AirIAD-8B |
- Release Training Scripts (Support for both Qwen3-VL-Instruct 4B and 8B versions).
- Release Testing & Evaluation Scripts for the MMAD benchmark.
- Open-source the training datasets (Warm-up SFT & RL data).
We would like to express our sincere gratitude to the following outstanding open-source projects and pioneering research works, which greatly inspired and facilitated the development of AirIAD:
- Training Frameworks: We thank the developers of verl and LLaMA-Factory for their robust and efficient training infrastructures.
- Benchmarks & Datasets: Special thanks to the creators of MMAD for providing comprehensive benchmarks that drive IAD research forward.
- Related MLLM & Reasoning Frameworks: We draw inspiration from excellent repositories including M3-AD, JUDO, IAD-R1, AgentIAD and AnomalyR1.