diff --git a/CHANGELOG.md b/CHANGELOG.md new file mode 100644 index 0000000..4b1b945 --- /dev/null +++ b/CHANGELOG.md @@ -0,0 +1,36 @@ +# Changelog + +Notable changes to the Mano-P project and its released models. Format loosely follows [Keep a Changelog](https://keepachangelog.com/). + +## [Unreleased] + +### Documentation + +- Added a 🚀 Quick Start section (EN/CN) with Homebrew install and cloud/local run examples +- Embedded the Video-SME-2, MIT1003 & SalECI, ETMD, and Online-Mind2Web benchmark charts above their data tables +- English hero banner now links to the in-page benchmark section instead of an external wiki +- Renamed `pics/MIT1003, SalEC.png` → `pics/MIT1003-SalECI.png`; removed unused `mano-action.jpeg` +- Contribution guidelines now state that `README.md` and `README_CN.md` must be kept in sync + +## [1.1] - 2026-06-01 + +### Models + +- Model naming standardized: released weights are now published as **Mano-CUA-4B-Thinking-1.1**, with an **MLX-8bit** variant for on-device inference on Apple Silicon +- Distribution expanded to Hugging Face, ModelScope CN, and ModelScope AI; model links added to the document header + +### Benchmarks + +- Refreshed all benchmark figures to model version 1.1 — **58.2% on OSWorld** (#1 among specialized GUI agent models), **41.7 NavEval** on WebRetriever Protocol I +- Added Instruct vs Thinking comparison on the 100-task execution suite: Mano-CUA-Thinking-4B 56.0% vs Mano-CUA-Instruct-4B 47.0% pass rate (cloud 83.0%, Qwen3-VL-Plus 39.0%) +- Added local (on-device) vs cloud vs Qwen3-VL task-execution comparison with per-tier and per-category breakdowns + +## [1.0] - 2026-03-25 + +### Added + +- Initial public release of the Mano-P project documentation +- Benchmark results: GUI grounding, OSWorld, WebRetriever, perception & cognition (Video-SME-2, MIT1003 & SalECI, ETMD), pruning (Online-Mind2Web / GSPruning), context learning (CL Bench) +- `mano-cua` CLI (Homebrew) and `mano-skill` (ClawHub) usage guides +- Cider inference-acceleration SDK documentation (W8A8 / W4A8 activation quantization for MLX) +- Bilingual README (English / 中文) diff --git a/README.md b/README.md index c4dd0e7..e52a33d 100644 --- a/README.md +++ b/README.md @@ -20,14 +20,14 @@ 中文 | English -**[📖 Overview](#-overview) | [🎯 Key Highlights](#-key-highlights) | [🎬 Use Cases](#-use-case-demonstrations) | [📊 Benchmark](#-benchmark-performance) | [🔧 Skills](#-skills) | [🤖 Models](#-models) | [⚡ Inference SDK](#-inference-sdk) | [⚗️ Approach](#-approach) | [🌟 Technical Advantages](#-technical-advantages) | [🔗 Applications](#-applications) | [📄 Citation](#-technical-papers--citation) | [❓ FAQ](#-faq)** +**[📖 Overview](#-overview) | [🎯 Key Highlights](#-key-highlights) | [🚀 Quick Start](#-quick-start) | [🎬 Use Cases](#-use-case-demonstrations) | [📊 Benchmark](#-benchmark-performance) | [🔧 Skills](#-skills) | [🤖 Models](#-models) | [⚡ Inference SDK](#-inference-sdk) | [⚗️ Approach](#-approach) | [🌟 Technical Advantages](#-technical-advantages) | [🔗 Applications](#-applications) | [📄 Citation](#-technical-papers--citation) | [❓ FAQ](#-faq)** ---
- + GUI Agent Grounding Benchmark
@@ -70,6 +70,31 @@ Mano-P builds upon the complete technical framework of the Mano project (see [Ma --- +## 🚀 Quick Start + +Get Mano-P running on your Mac in two minutes — see [🔧 Skills](#-skills) for full details: + +```bash +# Install the CLI via Homebrew +brew tap Mininglamp-AI/tap +brew install mano-cua + +# Run a task in cloud mode (default, no extra setup) +mano-cua run "Open Safari and search for Python" + +# Or run fully on-device (macOS Apple Silicon) — one-time setup: +mano-cua check +mano-cua install-sdk +mano-cua install-model + +# …then add --local: screenshots and task data never leave your device +mano-cua run "Open Safari and search for Python" --local +``` + +> **First run**: grant **Screen Recording** and **Accessibility** permissions to your terminal app in *System Settings → Privacy & Security*. + +--- + ## 🎬 Use Case Demonstrations ### Scenario 1: Mano-AFK Fully automated application construction @@ -154,6 +179,8 @@ This video demonstrates the combined capability of [Mano-AFK](#-applications) an #### Video-SME-2 +![Video-SME-2 results](./pics/Video-SME-2.png) + @@ -576,6 +603,8 @@ This video demonstrates the combined capability of [Mano-AFK](#-applications) an #### MIT1003 & SalECI +![MIT1003 & SalECI results](./pics/MIT1003-SalECI.png) +
@@ -713,6 +742,8 @@ This video demonstrates the combined capability of [Mano-AFK](#-applications) an #### ETMD +![ETMD results](./pics/ETMD.png) + ##### **Saliency Metrics**
@@ -822,6 +853,8 @@ _Avg. Tokens/img_ represents the average visual token retention rate per image; **GSPruning** is a novel token pruning method designed for Vision-Language Models to efficiently process high-resolution web interfaces by preserving global spatial structure through anchor points and identifying semantic outliers for critical UI elements. It achieves 2-3× throughput speedup with minimal performance loss, enabling more efficient autonomous web agents. +![Online-Mind2Web results](./pics/Online-Mind2Web.png) +
@@ -1670,6 +1703,8 @@ We welcome community contributions! If you want to contribute to the project: 4. Push to the branch (`git push origin feature/AmazingFeature`) 5. Open a Pull Request +> **Note**: `README.md` and `README_CN.md` are maintained in parallel — any change to one must be mirrored in the other. + ### Contribution Areas - 🐛 Bug fixes and issue reporting diff --git a/README_CN.md b/README_CN.md index 11b9c2a..4462bdc 100644 --- a/README_CN.md +++ b/README_CN.md @@ -20,7 +20,7 @@ 中文 | English -**[📖 项目概述](#-项目概述) | [🎯 核心亮点](#-核心亮点) | [🎬 应用场景](#-应用场景展示) | [📊 基准测试](#-基准测试性能) | [🔧 Skills](#-skills) | [🤖 模型](#-模型) | [⚡ 推理加速 SDK](#-推理加速-sdk) | [⚗️ 方法](#-方法) | [🌟 技术优势](#-技术优势) | [🔗 应用](#-应用) | [📄 论文引用](#-技术论文与引用) | [❓ FAQ](#-常见问题)** +**[📖 项目概述](#-项目概述) | [🎯 核心亮点](#-核心亮点) | [🚀 快速开始](#-快速开始) | [🎬 应用场景](#-应用场景展示) | [📊 基准测试](#-基准测试性能) | [🔧 Skills](#-skills) | [🤖 模型](#-模型) | [⚡ 推理加速 SDK](#-推理加速-sdk) | [⚗️ 方法](#-方法) | [🌟 技术优势](#-技术优势) | [🔗 应用](#-应用) | [📄 论文引用](#-技术论文与引用) | [❓ FAQ](#-常见问题)** @@ -69,6 +69,31 @@ Mano-P 基于完整的 Mano 项目技术体系（详见 [Mano Technical Report]( --- +## 🚀 快速开始 + +两分钟在 Mac 上把 Mano-P 跑起来——完整说明见 [🔧 Skills](#-skills)： + +```bash +# 通过 Homebrew 安装 CLI +brew tap Mininglamp-AI/tap +brew install mano-cua + +# 云端模式运行任务（默认，无需额外配置） +mano-cua run "打开 Safari 并搜索 Python" + +# 或完全在设备本地运行（macOS Apple Silicon）——首次需一次性安装： +mano-cua check +mano-cua install-sdk +mano-cua install-model + +# ……之后加 --local 即可：截图与任务数据不离开设备 +mano-cua run "打开 Safari 并搜索 Python" --local +``` + +> **首次运行**：请在 *系统设置 → 隐私与安全性* 中为你的终端 App 授予**屏幕录制**与**辅助功能**权限。 + +--- + ## 🎬 应用场景展示 ### 场景 1: Mano-AFK 全自动化应用构建 @@ -153,6 +178,8 @@ https://github.com/user-attachments/assets/5215d4eb-4e6f-4e03-b31b-dc8037c3794d #### Video-SME-2 +![Video-SME-2 评测结果](./pics/Video-SME-2.png) +
@@ -575,6 +602,8 @@ https://github.com/user-attachments/assets/5215d4eb-4e6f-4e03-b31b-dc8037c3794d #### MIT1003 & SalECI +![MIT1003 & SalECI 评测结果](./pics/MIT1003-SalECI.png) +
@@ -712,6 +741,8 @@ https://github.com/user-attachments/assets/5215d4eb-4e6f-4e03-b31b-dc8037c3794d #### ETMD +![ETMD 评测结果](./pics/ETMD.png) + ##### **Saliency Metrics**
@@ -821,6 +852,8 @@ _Avg. Tokens/img_ 表示每张图片的平均视觉 token 保留率；值越低 **GSPruning** 是一种新型的token剪枝方法，专为视觉语言模型设计，通过保留全局空间锚点维持网页结构骨架，并识别语义异常值来捕获关键UI元素，从而高效处理高分辨率网页界面。该方法在性能损失极小的情况下实现了2-3倍的吞吐量提升，为构建高效的自主网页智能体设立了新的技术标杆。 +![Online-Mind2Web 评测结果](./pics/Online-Mind2Web.png) +
@@ -1667,6 +1700,8 @@ brew tap Mininglamp-AI/tap && brew install mano-cua 4. 推送到分支 (`git push origin feature/AmazingFeature`) 5. 开启一个 Pull Request +> **注意**：`README.md` 与 `README_CN.md` 是并行维护的——修改任意一份必须同步修改另一份。 + ### 贡献方向 - 🐛 Bug修复和问题报告 diff --git a/pics/MIT1003, SalEC.png b/pics/MIT1003-SalECI.png similarity index 100% rename from pics/MIT1003, SalEC.png rename to pics/MIT1003-SalECI.png diff --git a/pics/mano-action.jpeg b/pics/mano-action.jpeg deleted file mode 100644 index e9defd9..0000000 Binary files a/pics/mano-action.jpeg and /dev/null differ