Qwen: Qwen3 VL 32B Instruct

Processes images and text simultaneously using a unified transformer architecture that fuses visual tokens from a vision encoder with text embeddings, enabling the model to answer questions about image content, describe visual scenes, and reason across visual and textual information in a single forward pass. The 32B parameter scale allows for nuanced spatial reasoning and semantic understanding of complex visual compositions.

Accepts video input (or sequences of frames) and performs temporal reasoning by processing multiple frames in context, understanding motion, scene changes, and temporal relationships between visual elements. The model maintains coherence across frames through attention mechanisms that track object persistence and state changes, enabling understanding of video narratives and dynamic visual events.

Analyzes document images (PDFs, scans, screenshots) to extract text, tables, and structured data with layout awareness. Uses visual understanding to identify table boundaries, column headers, and cell content, then outputs structured formats (JSON, CSV, Markdown) that preserve the original document structure. The model understands document semantics including headers, footers, and multi-column layouts.

Answers natural language questions about images by performing multi-step visual reasoning. The model decomposes complex questions into sub-questions, locates relevant visual regions, and chains reasoning steps together to arrive at answers. Supports both factual questions (what objects are present) and reasoning questions (why, how, what if) by leveraging the 32B parameter capacity for deeper inference.

Classifies images into semantic categories and generates descriptive tags by analyzing visual content. The model identifies objects, scenes, activities, and attributes present in images, then maps them to predefined or open-ended category systems. Supports both zero-shot classification (without training examples) and few-shot adaptation through in-context learning.

Executes complex, multi-step instructions that combine visual and textual inputs, following detailed specifications for output format, reasoning style, and content constraints. The model parses structured prompts (including system instructions, few-shot examples, and detailed task descriptions) and applies them consistently across multimodal inputs. Supports instruction-following patterns like chain-of-thought, role-playing, and format specifications.

Analyzes images for potentially harmful, inappropriate, or policy-violating content by identifying visual elements that may require moderation. The model detects violence, explicit content, hate symbols, misinformation indicators, and other safety-relevant visual patterns. Provides confidence scores and detailed explanations for moderation decisions, enabling human-in-the-loop review workflows.

Understands spatial relationships, object positions, and scene composition by analyzing visual layouts. The model identifies foreground/background relationships, depth cues, spatial arrangements, and geometric relationships between objects. Supports queries about relative positions, occlusion, perspective, and scene structure, enabling applications that require spatial reasoning beyond simple object detection.

Recognizes and extracts text from images while understanding context and language semantics. Beyond character-level OCR, the model comprehends text meaning, identifies text language, handles multiple scripts, and understands text in context (e.g., captions, labels, handwriting). Supports text extraction from complex layouts including rotated text, overlapping text, and variable font sizes.