第一项始终是图钉的来源。总体可信度是各份资料对上方所用开始和结束时间支持程度的加权平均;资料每比最新的一份旧 180 天,权重减半。
模型卡的项目更新写道“2025/08/11: We released GLM-4.5V”;OpenRouter[6]于2025年8月11日上架z-ai/glm-4.5v,arXiv[5]论文的GLM-4.5V修订版于8月13日跟进,而Hugging Face权重则在UTC时间8月10日上传。
Z.ai developer docs: GLM-4.5V visual reasoning model on an MoE architecture with 106B total and 12B active parameters, based on GLM-4.5-Air; Thinking Mode, web-page coding, grounding, GUI agents, long-document and video understanding; $0.6 and $1.8 per million input and output tokens, 16K max output; 42 benchmarks.[1]
SiliconFlow launch post for its hosted GLM-4.5V: state of the art on 42 vision-language benchmarks, 66K-token context, 3D convolution for video, bicubic interpolation, 3D-RoPE, and three-stage training with RLVR and RLHF; hosted at $0.14 and $0.86 per million tokens. Page date is internally inconsistent, so it is not used for dating.
Labellerr explainer (15 Sep 2025): GLM-4.5V is the vision-language member of the GLM-4.5 family, built on GLM-4.5-Air, with 3D-RoPE for spatial relationships, image and video reasoning, GUI-agent tasks, chart parsing and precise grounding; source of the architecture figure.
GLM-V paper (v3 of 13 Aug 2025 added GLM-4.5V): evaluation on 42 public benchmarks, state of the art on nearly all tasks among similar-size open models and competitive with or better than Gemini-2.5-Flash on some tasks including coding and GUI agents; GLM-4.1V-9B-Thinking beats Qwen2.5-VL-72B on 29 benchmarks.[1]
有遗漏或错误吗?用你自己的话说明:能佐证此图钉的链接、不同的开始或结束日期及理由,或缺失、有误的信息。AI 会对照此图钉的来源进行核实,搜索更好的来源,并添加任何支持你说法的页面。图钉自身的来源仍然最重要。AI 也会查看图片:显示的是别的东西或显示效果差的图片会被移到后面或替换。
原标题: GLM-4.5V Released
Z.ai在docs.z.ai上发布的GLM-4.5V图表中的部分行;“*”表示Z.ai自行复现的结果,“/”表示未报告[2].
| 基准 | GLM-4.5V (106B-A12B, 思考) | Step-3 (321B-A38B, 思考) | Qwen2.5-VL 72B (非思考) | GLM-4.1V 9B (思考) | Kimi-VL-2506 (16B-A3B, 思考) | Gemma-3 27B (非思考) |
|---|---|---|---|---|---|---|
| MMBench v1.1 | 88.2 | 81.1* | 88.0 | 85.8 | 84.4 | 80.1* |
| MMStar | 75.3 | 69.0* | 70.8 | 72.9 | 70.4 | 60.0* |
| MMMU (val) | 75.4 | 74.2 | 70.2 | 68.0 | 64.0 | 62.0* |
| MMMU Pro | 65.2 | 58.6 | 51.1 | 57.1 | 46.3 | 37.4* |
| MathVista | 84.6 | 79.2* | 74.8 | 80.7 | 80.1 | 64.3* |
| MathVision | 65.6 | 64.8 | 38.1 | 54.4 | 54.4* | 39.8* |
| AI2D | 88.1 | 83.7* | 87.6* | 87.9 | 81.9* | 80.2* |
| MMLongBench-Doc | 44.7 | 31.8* | 35.2* | 42.4 | 42.1 | 28.4* |
| OCRBench | 86.5 | 83.7 | 85.1* | 84.2 | 86.9 | 75.9* |
| RefCOCO-avg (val) | 91.3 | 20.2* | 90.3 | 85.3 | 33.6* | 2.4* |
| OSWorld | 35.8 | / | 8.8 | 14.9 | 8.2 | 6.2* |
| AndroidWorld | 57.0 | / | 35.0 | 41.7 | / | 4.4* |
| WebVoyagerSom | 84.4 | / | 40.4* | 69.0 | / | 34.8* |
| Design2Code | 82.2 | 34.1 | 41.9* | 64.7 | 38.8 | 16.1 |
| VideoMME (无字幕) | 74.6 | / | 73.3 | 68.2 | 67.8 | 58.9* |
| VideoMMMU | 72.4 | / | 60.2 | 61.0 | 65.2 | 54.5* |