Qwen3-VL-8B-Instruct

Qwen3-VL-8B-Instruct is a multimodal model with 8 billion parameters, offering an optimal balance between performance and deployment efficiency. Built upon the Qwen3-8B language model with an integrated Vision Transformer-based visual encoder, it provides seamless understanding of text, images, and videos. Thanks to architectural innovations—Interleaved-MRoPE, DeepStack, and Text-Timestamp Alignment—the model demonstrates superior multimodal comprehension, surpassing its predecessor, Qwen2.5-VL-7B, across all key accuracy metrics, while achieving a 15-60% increase in token generation speed and a 20-40% reduction in response latency.

A key feature of the model is its native support for a 256K token context window, expandable up to 1 million tokens. This enables the processing of entire books, multi-hour videos, and complex multi-page documents with full context retention. Enhanced OCR capabilities supporting 32 languages (up from 19 in the previous version) and robustness to challenging capture conditions make Qwen3-VL-8B-Instruct an ideal solution for intelligent document processing. The model accurately recognizes text under low light, blur, and tilt conditions, handles rare and ancient characters, and understands complex long-document structures. On the DocVQA benchmark, it shows a significant advantage due to its improved document structure parsing.

The model was trained on a significantly enriched multimodal corpus, ensuring nearly complete coverage of real-world object categories (faces, natural landscapes, products, and interfaces). In this context, its visual agent capabilities are particularly noteworthy: Qwen3-VL-8B-Instruct can recognize GUI elements (buttons, input fields, menus), understand their functions, and execute complex action sequences on PCs and mobile devices. It generates functional HTML/CSS/JavaScript code and Draw.io diagrams from images, significantly accelerating interface prototyping. Advanced spatial perception with support for 2D and 3D object localization opens up possibilities for applications in robotic vision and embodied AI.

Announce Date: 15.10.2025
Parameters: 9B
Context: 263K
Layers: 36
Attention Type: Full Attention
Developer: Qwen
Transformers Version: 4.57.0.dev0
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore Qwen3-VL-8B-Instruct capabilities. You can obtain an API access token on the token management page after registration and verification.

An endpoint is a server with pre-installed software that automatically downloads neural network weights and launches an OpenAI-compatible API to handle incoming requests.

A private endpoint is created within the user's project, belongs exclusively to the user, and is billed per server configuration in accordance with the infrastructure pricing.

We recommend creating private endpoints when you need to:

• achieve maximum endpoint performance,
• access the full context window for processing large requests and conversations,
• ensure maximum security by processing data in a dedicated, isolated environment.

Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.