Qwen3-VL

Qwen3-VL is Alibaba's flagship vision-language series, launched September 2025 with the 235B-A22B mixture-of-experts at the top and smaller 2B to 32B variants following. The flagship activates 22B parameters per token and carries 256K native context, expandable toward 1M tokens, which Alibaba uses to demonstrate analysis of two-hour videos with near-perfect frame retrieval in needle-in-a-haystack tests. It covers 2D and 3D spatial understanding, document VQA, chart and table reasoning, GUI interaction, and code generation from screenshots, with text-timestamp alignment for precise event localization in video. Instruct and Thinking variants shipped together. Alibaba reports wins over Gemini 2.5 Pro on several vision benchmarks; treat these as team-reported, since very high scores on imperfect benchmarks deserve skepticism. Everything is Apache 2.0 on Hugging Face, and the technical report is on arXiv. The 235B size is strictly server-side, but the 8B and smaller variants are practical for single-GPU work.

Team-reported: beats Gemini 2.5 Pro on several vision benchmarks; 100% 30-min video needle retrieval, 99.5% at 2 hours