shuai bai commited on
Commit
a85c315
·
verified ·
1 Parent(s): cbaeae4

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +32 -12
README.md CHANGED
@@ -5,28 +5,48 @@ license: apache-2.0
5
  <img alt="Chat" src="https://img.shields.io/badge/%F0%9F%92%9C%EF%B8%8F%20Qwen%20Chat%20-536af5" style="display: inline-block; vertical-align: middle;"/>
6
  </a>
7
 
8
- ## Introduction
9
- The goal of Qwen3-VL is not just to “see” images or videos — but to truly understand the world, interpret events, and take action. To achieve this, we’ve systematically upgraded key capabilities, pushing visual models from simple “perception” toward deeper “cognition,” and from basic “recognition” to advanced “reasoning and execution.”
10
- ![](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3-VL/qwen3vl_arc.jpg)
11
 
 
12
 
13
- **Key Highlights:**
14
 
15
- - **Visual Agent Capabilities**: Qwen3-VL can operate computer and mobile interfaces recognize GUI elements, understand button functions, call tools, and complete tasks. It achieves top global performance on benchmarks like OS World, and using tools significantly improves its performance on fine-grained perception tasks.
16
 
17
- - **Superior Pure-text Performance**: Qwen3-VL employs early-stage joint pretraining of text and visual modalities, continuously strengthening its language capabilities. It ultimately achieves pure-text task performance comparable to Qwen3-235B-A22B-2507 — the flagship pure-text model — making it a truly “text-grounded, multimodal powerhouse” for the next generation of vision-language models.
18
 
19
- - **Greatly Improved Visual Coding**: It can now generate code from images or videos — for example, turning a design mockup into Draw.io, HTML, CSS, or JavaScript code — making “what you see is what you get” visual programming a reality.
20
 
21
- - **Much Better Spatial Understanding**: 2D grounding is strengthened and the coordinate is transformed from absolute to relative, enabling more robust estimation of object locations, viewpoint (camera pose) changes, and occlusion structure. The model also supports 3D grounding, providing a foundation for complex spatial reasoning and embodied AI applications.
22
 
23
- - **Long Context & Long Video Understanding**: Natively support 256K tokens of context, expandable up to 1 million tokens. This means you can input hundreds of pages of technical documentation, entire textbooks, and even two‑hour videos of meetings or lectures — and achieves reliable long‑context retention and precise retrieval, including second‑level timestamp localization in video.
24
 
25
- - **Stronger Multimodal Reasoning (Thinking Version)**: The Thinking model is optimized for STEM reasoning. On complex problems, it attends to fine‑grained cues, performs step‑by‑step decomposition, analyzes causal dependencies, and produces logically consistent, evidence‑grounded solutions. It achieves state‑of‑the‑art results on benchmarks including MathVision, MMMU, and MathVista.
26
 
27
- - **Upgraded Visual Perception & Recognition**: Improvements to the quality and diversity of the pretraining corpus have expanded the model’s recognition ability to a broader range of objects and entities, including celebrities, anime characters, consumer products, landmarks, and flora/fauna, — covering both everyday life and professional “recognize anything” needs.
28
 
29
- - **Better OCR Across More Languages & Complex Scenes**: OCR now supports 32 languages (up from 10), covering more countries and regions. It demonstrates greater robustness under challenging real-world conditions like poor lighting, blur, or tilted text. Recognition accuracy for rare characters, ancient scripts, and technical terms has also improved significantly. Its ability to understand long documents and reconstruct fine structures is further enhanced.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
30
 
31
  This is the weight repository for Qwen3-VL-235B-A22B-Instruct.
32
 
 
5
  <img alt="Chat" src="https://img.shields.io/badge/%F0%9F%92%9C%EF%B8%8F%20Qwen%20Chat%20-536af5" style="display: inline-block; vertical-align: middle;"/>
6
  </a>
7
 
 
 
 
8
 
9
+ # Qwen3-VL-235B-A22B-Instruct
10
 
 
11
 
12
+ Meet Qwen3-VL — the most powerful vision-language model in the Qwen series to date.
13
 
14
+ This generation delivers comprehensive upgrades across the board: superior text understanding & generation, deeper visual perception & reasoning, extended context length, enhanced spatial and video dynamics comprehension, and stronger agent interaction capabilities.
15
 
16
+ Available in Dense and MoE architectures that scale from edge to cloud, with Instruct and reasoning‑enhanced Thinking editions for flexible, on‑demand deployment.
17
 
 
18
 
19
+ #### Key Enhancements:
20
 
21
+ * **Visual Agent**: Operates PC/mobile GUIs—recognizes elements, understands functions, invokes tools, completes tasks.
22
 
23
+ * **Visual Coding Boost**: Generates Draw.io/HTML/CSS/JS from images/videos.
24
 
25
+ * **Advanced Spatial Perception**: Judges object positions, viewpoints, and occlusions; provides stronger 2D grounding and enables 3D grounding for spatial reasoning and embodied AI.
26
+
27
+ * **Long Context & Video Understanding**: Native 256K context, expandable to 1M; handles books and hours-long video with full recall and second-level indexing.
28
+
29
+ * **Enhanced Multimodal Reasoning**: Excels in STEM/Math—causal analysis and logical, evidence-based answers.
30
+
31
+ * **Upgraded Visual Recognition**: Broader, higher-quality pretraining is able to “recognize everything”—celebrities, anime, products, landmarks, flora/fauna, etc.
32
+
33
+ * **Expanded OCR**: Supports 32 languages (up from 19); robust in low light, blur, and tilt; better with rare/ancient characters and jargon; improved long-document structure parsing.
34
+
35
+ * **Text Understanding on par with pure LLMs**: Seamless text–vision fusion for lossless, unified comprehension.
36
+
37
+
38
+ #### Model Architecture Updates:
39
+
40
+ <p align="center">
41
+ <img src="https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3-VL/qwen3vl_arc.jpg" width="80%"/>
42
+ <p>
43
+
44
+
45
+ 1. **Interleaved-MRoPE**: Full‑frequency allocation over time, width, and height via robust positional embeddings, enhancing long‑horizon video reasoning.
46
+
47
+ 2. **DeepStack**: Fuses multi‑level ViT features to capture fine‑grained details and sharpen image–text alignment.
48
+
49
+ 3. **Text–Timestamp Alignment:** Moves beyond T‑RoPE to precise, timestamp‑grounded event localization for stronger video temporal modeling.
50
 
51
  This is the weight repository for Qwen3-VL-235B-A22B-Instruct.
52