Qwen3-VL 2B Instruct

Qwen3-VL-2B-Instruct is a vision-language model (VLM) in the Qwen3 series, designed for tasks requiring simultaneous understanding and generation of images, video, and text. The model delivers comprehensive visual perception and reasoning capabilities: it can operate PC and mobile graphical user interfaces as an agent, generate structured visual code (Draw.io, HTML, CSS, JS) from images or videos, and perform advanced spatial reasoning including object localization and 3D grounding. It supports a native 256K context window (expandable to 1M), enabling processing of long-form video with second-level event localization. OCR is expanded to 32 languages and performs robustly under low light, blur, and tilt. The model also excels in STEM and mathematical reasoning, providing evidence-based answers, and its text understanding is on par with pure LLMs through seamless vision-text fusion.

Architecture

The model incorporates three key architectural innovations:

• Interleaved-MRoPE – full-frequency positional embeddings over time, width, and height for improved video reasoning.
• DeepStack – fuses multi-level ViT features to capture fine-grained details and sharpen image-text alignment.
• Text–Timestamp Alignment – enables precise, timestamp-grounded event localization for temporal video modeling.

Performance

Evaluated on standard multimodal and pure-text benchmarks, Qwen3-VL-2B-Instruct demonstrates competitive results within its size class.

The model is released under the Apache 2.0 license. It supports a thinking budget mechanism for adaptive inference resource allocation, and its training leverages knowledge from larger flagship models to achieve strong performance with minimal computational overhead.