Qwen3-VL-30B-A3B-Instruct and Thinking (huggingface.co)

Alibaba’s Qwen team has released Qwen3‑VL, the most capable vision‑language model in the Qwen family — now available as the 30B "Qwen3‑VL‑30B‑A3B‑Thinking" weights (Dense and MoE variants) and offered in both Instruct and reasoning‑focused "Thinking" editions. The model is pitched as a unified text+vision system with LLM‑level text understanding, stronger multimodal reasoning (notably STEM/math), GUI‑capable visual agent behavior, and visual coding features that generate Draw.io/HTML/CSS/JS from images/videos. It also improves recognition scope (celebrities, anime, products, flora/fauna), expands OCR from 19 to 32 languages, and supports robust reading under blur/tilt and rare scripts — enabling applications from embodied agents and GUI automation to hours‑long video summarization and large‑scale document understanding. Technically, Qwen3‑VL brings several architecture and modeling advances: native 256K token context (scalable to 1M with second‑level indexing) for books and long videos, Interleaved‑MRoPE positional embeddings for full‑frequency time/space allocation in long‑horizon video reasoning, DeepStack fusion of multi‑level ViT features for finer image–text alignment, and Text‑Timestamp Alignment that improves temporal grounding beyond T‑RoPE. The repo is integrated with Hugging Face/ModelScope (example code available) and recommends memory/latency optimizations like flash_attention_2 and bfloat16; MoE vs Dense choices let teams trade compute for edge‑to‑cloud deployment flexibility. Overall, Qwen3‑VL pushes VLMs toward persistent, temporally precise multimodal agents and large‑context video/document understanding.