Qwen3-VL-8B-Instruct

Qwen3-VL-8B-Instruct is an open-source dense vision-language model released by Alibaba's Qwen team on October 15, 2025, under the Apache 2.0 license. Built on the Qwen3-8B language backbone with a native-resolution Vision Transformer encoder, it accepts interleaved text, image, and video inputs and produces text outputs. Key architectural innovations include DeepStack multi-level visual feature fusion, Interleaved-MRoPE for spatial-temporal positional encoding, and text-based timestamp alignment for video understanding. The model natively supports a 256K-token context window (expandable to 1M in some deployments) and 32-language OCR.

On benchmarks, Qwen3-VL-8B-Instruct delivers strong multimodal performance for its size: MMMU val ~69.6, DocVQA ~96.1, OCRBench ~896, and MMBench-EN ~85.3 per the official Qwen3-VL technical report. Pure-text capability is largely preserved versus Qwen3-8B, scoring ~80.7 on MMLU (Artificial Analysis), ~81.5 on GSM8K, and ~68.0 on HumanEval. It ranks competitively among open 8B VLMs and is widely deployed via Hugging Face, ModelScope, Ollama, vLLM, and Alibaba Cloud API.

The model targets edge-to-cloud multimodal applications including document OCR and parsing, visual question answering, GUI/visual agent control, spatial grounding, chart extraction, and long-video reasoning. Both Instruct and Thinking variants are available; the Thinking edition adds chain-of-thought reasoning for harder STEM and visual-math tasks at higher inference cost.