Qwen3 VL 4B Instruct
Qwen/Qwen3-VL-4B-Instruct
published Oct 2025 · updated Oct 2025
Qwen3 VL 4B Instruct is a vision-language model that excels in visual reasoning, long-context video understanding, and agent tasks, with dynamic thinking mode.
specs
| Task | Vision-Language (VLM) |
| Architecture | Dense Vision-Language Transformer with Interleaved-MRoPE and DeepStack |
| Parameters | 4 billion |
| License | Apache 2.0 |
about this model
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.
best for
- ·GUI automation and agent tasks
- ·Visual coding from images/videos
- ·Long-document OCR and parsing
- ·Video event localization and comprehension
FAQ
It excels at visual reasoning, long-context video understanding, visual agent tasks, and OCR in 32 languages.
It accepts images, videos, and text as input; outputs generated text. Use chat template with roles.
Use the gigarouter OpenAI-compatible endpoint with an API key; send messages with image and text content.
Native 256K tokens, expandable to 1M tokens.
Apache 2.0 license, allowing free use, modification, and distribution.
# rerank documents by relevance; billed per document curl https://gigarouter.ai/v1/rerank \ -H "Authorization: Bearer $GR_KEY" \ -d '{"model":"Qwen/Qwen3-VL-4B-Instruct","query":"capital of France", "documents":["Paris is the capital of France.","Bananas are yellow."]}'
try it live
runs the real hosted model on a shared demo allowance · get your own key + $25 free →