Qwen3-VL-32B-Instruct

Qwen3-VL-32B-Instruct is the flagship dense model of the Qwen3-VL family with 32 billion parameters, demonstrating top-tier results on a wide range of multimodal benchmarks. The model achieves performance comparable to leading proprietary systems like GPT-5 and Claude 4 Sonnet, while surpassing other open-source models. The model's architectural foundation is built upon three key innovations. Interleaved-MRoPE implements full-spectrum positional information distribution across the temporal axis and spatial dimensions (width/height), providing robust positional encoding for efficient video analysis. DeepStack integrates Vision Transformer and text features, significantly enhancing their alignment. Text-Timestamp Alignment ensures precise timestamp anchoring of events for temporal video modeling. This architecture enables the model to process visual and textual information simultaneously.

The model natively supports a 256K token context window with the capability to extend up to 1M tokens, allowing it to process extensive documents, entire books, or long videos with full memory recall and accurate information extraction. Visual perception is enhanced through large-scale, high-quality training—the model can recognize celebrities, anime characters, products, locations, flowers, animals, and other objects. The OCR system supports 32 languages (compared to 19 in previous versions), including Greek, Hebrew, Hindi, Romanian, Thai, and many others, with improved robustness under challenging conditions like poor lighting, blur, and skew. Enhanced accuracy in recognizing rare symbols, ancient scripts, and technical terminology expands the model's applicability in specialized domains. The model can function as a visual agent, recognizing user interface elements on computers and mobile devices, understanding their functions, capable of invoking tools, and confidently executing multi-step tasks.

Qwen3-VL-32B-Instruct is optimal for enterprise-grade visual analysis applications with high accuracy requirements. Automation of document, contract, and financial report processing. Advanced visual agents for UI testing automation, RPA scenarios, and managing complex GUI systems. Large-scale video analytics for security, retail, and media with temporal event localization, indexing, and search. Multimodal big data analytics, processing visual and textual data in a unified pipeline. Code generation from designs to accelerate front-end application development.

Announce Date: 22.10.2025
Parameters: 33B
Context: 263K
Layers: 64
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-32B-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.