Qwen3 VL 4B Instruct

Qwen3-VL-4B-Instruct is a vision-language model (VLM) that integrates advanced visual perception, reasoning, and agent interaction capabilities with native 256K context length (expandable to 1M) for processing images, videos, and long-form text.

Capabilities

• Visual agent functionality: operates PC and mobile GUIs by recognizing elements, understanding functions, invoking tools, and completing tasks.
• Visual code generation: produces Draw.io, HTML, CSS, and JS from image or video inputs.
• Advanced spatial perception: judges object positions, viewpoints, and occlusions; supports 2D grounding and 3D grounding for spatial reasoning and embodied AI.
• Long-context video understanding: handles hours-long video with full recall and second-level event indexing.
• Upgraded visual recognition: identifies celebrities, anime, products, landmarks, flora/fauna, and other categories.
• Expanded OCR: supports 32 languages (up from 19), performs robustly in low light, blur, and tilt conditions, and handles rare/ancient characters and jargon.
• Text understanding on par with pure LLMs, achieved through seamless text–vision fusion.

Architecture

• Interleaved-MRoPE: full-frequency allocation over time, width, and height for robust positional embeddings, enhancing 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 in the figures below.

This model is part of the Qwen3-VL family, which incorporates thinking and non-thinking modes in a unified framework, and supports adaptive reasoning budgets. Qwen3-VL-4B-Instruct is available as a hosted, OpenAI-compatible API on gigarouter.