Qwen3-VL

2025-09-23 · Publisher : Alibaba · Domain : Artificial Intelligence

Alibaba Qwen releases Qwen3-VL-235B-A22B in Instruct and Thinking versions under Apache 2.0, followed by a family of dense and MoE sizes (2B to 32B, 30B-A3B); native 256,000-token context and a visual agent.

### Context Qwen2.5-VL (January 2025) was the family's last major open vision-language model. Qwen3-VL builds on Qwen3 and is compared with Gemini 2.5 Pro, GPT-5 and Claude Opus 4.1 according to the technical report. ### What the model brings - Visual agent: control of PC and mobile interfaces (element recognition, tool calling, task execution). - Visual code generation (Draw.io, HTML, CSS, JavaScript) from images and videos. - Enhanced spatial perception: positions, viewpoints, occlusions, 2D and 3D grounding. - Native context of 256,000 tokens, extensible to 1 million, and understanding of videos lasting several hours with second-level indexing. - OCR extended to 32 languages (versus 19) and text performance on par with pure language models. - Three architecture innovations: Interleaved-MRoPE (allocation of all frequencies to time, width and height), DeepStack (fusion of multi-level visual features) and text-timestamp alignment for video. ### Architecture and specifications Technical report (arXiv 2511.21631): four dense models (2B, 4B, 8B, 32B) and two MoEs (30B-A3B, 235B-A22B), trained up to 256,000 tokens of context in four phases (alignment, then contexts of 8 K, 32 K and 256 K). Staggered release: 235B-A22B on September 23, 2025, 30B-A3B on October 4, 4B and 8B on October 15, 2B and 32B on October 21. Apache 2.0 license. Weights on Hugging Face and ModelScope, FP8 versions available. Knowledge cutoff: not disclosed. ### Cost and access | Item | Value | |---|---| | Weights | Free, Apache 2.0 | | qwen3-vl-235b-a22b Instruct API (current pricing consulted) | $0.4 input / $1.6 output per million tokens | | qwen3-vl-235b-a22b Thinking API (current pricing consulted) | $0.4 input / $4 output | ### Benchmarks The multimodal tables in the post and the card are images. The technical report publishes, in text, the comparison on the text capabilities of Qwen3-VL-235B-A22B-Thinking (vendor measurements). | Benchmark | OpenAI o3 | Claude Opus 4 (thinking) | Qwen3-VL-235B-A22B-Thinking | |---|---|---|---| | MMLU-Pro | 85.9 | not published | 83.8 | | GPQA | 83.3 | 79.6 | 77.1 | | AIME-25 | 88.9 | 75.5 | 89.7 | | LiveCodeBench v6 | 58.6 | 48.9 | 70.1 | | BFCL-v3 | 72.4 | 61.8 | 71.8 | | Arena-Hard V2 | 80.8 | 59.1 | 74.8 | Source: Qwen3-VL technical report (arXiv 2511.21631), table 3. On visual tasks, the report states that the Thinking version scores 78.7 on MMStar, the best score among the reasoning models compared, with Gemini 2.5 Pro Thinking being the best overall. ### Versus closed models On text capabilities, the Thinking version outperforms o3 on AIME-25 (89.7 versus 88.9) and LiveCodeBench v6 (70.1 versus 58.6) and stays behind on GPQA (77.1 versus 83.3) and Arena-Hard V2 (74.8 versus 80.8). On vision, the report acknowledges that Gemini 2.5 Pro in thinking mode delivers the best overall results, with Qwen3-VL-235B-A22B-Thinking being "not far behind" according to the authors, and describes gains on HallusionBench against Gemini 2.5 Pro, GPT-5 and Claude Opus 4.1. The detailed visual values were not transcribed for lack of a text table. The 32-billion-parameter model is claimed to be above Gemini 2.5 Flash and GPT-5 mini on most of the report's tests. For enterprise use, this opens up document, video and screen analysis with self-hosting. ### Key takeaways Qwen3-VL is Qwen's vision-language milestone for 2025: a complete family, from the 2-billion-parameter model to the 235-billion-parameter MoE, with a visual agent and long context. Fine-grained visual comparisons should be consulted in the technical report.

Tags : LLM, Open Source, Alibaba, Qwen, MoE, Vision, Apache 2.0, Multimodal