Qwen3-VL-4B-Instruct

Qwen3-VL-4B-Instruct is a dense vision-language model (VLM) built for multimodal understanding, visual grounding, and agentic tasks. It is hosted as a managed, OpenAI-compatible API on gigarouter, allowing developers to integrate vision-language capabilities without managing infrastructure.

Key Capabilities

• Visual Agent: Operates PC and mobile GUIs by recognizing elements, understanding functions, and invoking tools to complete tasks.
• Visual Coding: Generates Draw.io diagrams, HTML, CSS, and JS directly from images or videos.
• Spatial Perception: Judges object positions, viewpoints, and occlusions. Supports 2D grounding and enables 3D grounding for spatial reasoning and embodied AI.
• Long Context & Video Understanding: Native 256K context, expandable to 1M tokens. Handles books and hours-long video with full recall and second-level indexing.
• Multimodal Reasoning: Excels in STEM and math with causal analysis and evidence-based answers.
• Visual Recognition: Broad-covering pretraining recognizes celebrities, anime, products, landmarks, flora, fauna, and more.
• Expanded OCR: Supports 32 languages (up from 19). Robust under low light, blur, and tilt. Handles rare and ancient characters, jargon, and improves long-document structure parsing.
• Text Understanding: Seamless text–vision fusion yields comprehension on par with pure large language models.

Architecture Highlights

• Interleaved-MRoPE: Full-frequency allocation over time, width, and height for enhanced long-horizon video reasoning.
• DeepStack: Fuses multi-level ViT features to capture fine-grained details and sharpen image–text alignment.
• Text–Timestamp Alignment: Precise, timestamp-grounded event localization for stronger video temporal modeling.

Performance

Multimodal and pure text benchmark results are shown below.