skip to content
gigarouter gigarouter
models / vision-language · coming soon

Qwen3-VL 2B Instruct

Qwen/Qwen3-VL-2B-Instruct

published Oct 2025 · updated Oct 2025

Qwen3-VL 2B Instruct is a vision-language model that delivers advanced text understanding, visual perception, reasoning, and agent capabilities for images, videos, and documents.

est. price
~$0.626
/ 1k images · estimated, set at launch
API providers
0
downloads / mo
2.1M
license
apache-2.0

specs

TaskVision-Language Understanding & Generation
ArchitectureDense Transformer with Interleaved-MRoPE, DeepStack, and Text–Timestamp Alignment
Parameters2B
LicenseApache 2.0

about this model

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.
Architecture diagram showing Interleaved-MRoPE, DeepStack, and Text-Timestamp Alignment

Performance

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

Multimodal performance chart Pure text performance chart

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.

best for

FAQ

What is the context length supported by Qwen3-VL 2B Instruct?

It supports native 256K tokens, expandable to 1M tokens for long documents and hours-long video.

How does Qwen3-VL 2B Instruct compare to larger Qwen3-VL models?

It is the smallest dense variant in the Qwen3-VL family, optimized for edge and lightweight deployments while still offering strong vision-language capabilities.

What is the thinking budget mechanism?

It allows users to allocate computational resources adaptively during inference, balancing latency and performance based on task complexity.

How many languages does the model support for OCR?

OCR supports 32 languages; the underlying Qwen3 LLM supports 119 languages for text.

How can I call this model via the gigarouter API?

Use the OpenAI-compatible endpoint with your gigarouter API key, sending image and text inputs in the standard chat completion format.

not yet live

We're benchmarking and onboarding Qwen3-VL 2B Instruct as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.

related vision-language models

compare all →