Xuanwu VL-2B: An Industrial-Grade Multimodal Foundation Model for Content Ecosystems

Xuanwu VL-2B is an industrial-grade multimodal foundation model for content ecosystems. It adopts a compact architecture of InternViT-300M + MLP + Qwen3 1.7B (about 2B parameters). Through an iterative data filtering mechanism and three-stage progressive training, it achieves a balance between business alignment, visual perception, and general capabilities. Its recall rate in adversarial OCR scenarios reaches 82.82%, surpassing Gemini-2.5-Pro; the average recall rate for business audit tasks is 94.38%; its general multimodal capabilities on the OpenCompass benchmark are superior to similar models, balancing deployment cost and efficiency.

In recent years, multimodal large language models have performed well on academic benchmarks, but when deployed to content ecosystems (such as content audit and ad recognition), they face three major challenges: 1. Fine-grained visual perception requirements (recognizing tiny details, text, and implicit symbols); 2. Robustness to adversarial samples (dealing with malicious bypass methods like image distortion and text occlusion); 3. Long-tail distribution problem (violating content types are diverse and rare). These cause academic high-score models to have reduced generalization ability in industrial scenarios and easily forget general capabilities.

Xuanwu VL-2B adopts a three-component architecture:

1. Visual Encoder: InternViT-300M (lightweight, balancing fine-grained perception and computational overhead);
2. Projection Layer: MLP (connects visual and language feature spaces to ensure semantic retention);
3. Language Model: Qwen3 1.7B (optimized for Chinese, supports large-scale deployment with efficient inference). The overall parameter size is about 2B, achieving "small size with great power".

The model uses three-stage training:

1. Pre-training: Establishes cross-modal basic capabilities using large-scale general multimodal data;
2. Mid-training: The core innovation is the iterative data filtering mechanism, which identifies and removes low-quality samples through model feedback and supplements high-quality data;
3. Post-training: Aligns with scenarios using business datasets (audit samples, adversarial samples), and consolidates robustness through adversarial training and curriculum learning.

Evaluation verifies the model's effectiveness:

1. General Capabilities: An average score of 67.90 on the OpenCompass benchmark, superior to InternVL3.5 2B's 64.27;
2. Business Audit: An average recall rate of 94.38% across 7 tasks, effectively capturing violating content;
3. Adversarial OCR: A weighted recall rate of 82.82%, surpassing Gemini-2.5-Pro's 76.72%, proving that lightweight models can outperform large models in specific domains.

Xuanwu VL-2B is suitable for industrial deployment: 2B parameters can run on a single consumer-grade GPU or high-performance CPU, reducing hardware costs; at the same time, high recall rate (reducing missed detection risks), adversarial robustness (resisting malicious bypass), and general capability retention (adapting to business changes) form a reliable content audit infrastructure.

Insights from Xuanwu VL-2B:

1. Data quality first: The iterative filtering mechanism purifies training data to improve reliability;
2. Progressive training balance: Phased training avoids catastrophic forgetting and balances professionalism and generality;
3. Targeted architecture: Carefully selected components allow lightweight models to outperform large models. These experiences provide references for industrial AI system development.