Mano-P 1.0

GUI-Aware Agent Model for Edge Devices

Private AI

中文 | English

📖 Overview

Mano-P: "Mano" means "hand" in Spanish, and "P" stands for Private. We believe that both individuals and organizations can create their own Private AI, and a bright future of human-machine collaboration is on the horizon.

Mano-P is a GUI-VLA agent project designed specifically for edge devices. It serves both as an open-source project and a hardware product solution. As an open-source project, Mano-P is being released in a phased, progressive manner, targeting three distinct groups of developers. In the first phase, we will open-source the Mano-CUA Skills. This phase is aimed at Agent enthusiasts—such as users of OpenClaw or Claude Code—enabling them to leverage the capabilities of Mano-CUA Skills to construct more intelligent CUA task workflows and overcome the bottlenecks associated with human intervention. In the second phase, we will open-source the local-side models and SDK components of Mano-CUA. This phase targets developers with high security requirements, allowing them to directly utilize GUI-VLA models capable of running inference locally on a Mac mini to build their own custom Skills, Tools, and more; crucially, all your CUA operations will be executed entirely on your local Mac mini and will not be uploaded to external servers. In the third phase, we will open-source the training methodologies and the pruning and quantization techniques used for the Mano-P models. This phase is designed for developers with specific model training needs, empowering them to apply our training methods to create their own on-device GUI-VLA models tailored to their unique requirements.

Regarding our GUI-VLA models—which are capable of running inference directly on Mac mini and MacBook devices—we currently support two deployment methods: First, direct deployment on Mac mini or MacBook models equipped with an M4 chip and 32GB or more of RAM; and second, deployment utilizing a compute stick connected via a USB 4.0 port or higher. We will be releasing detailed instructions for both deployment methods in the near future, and we plan to expand our support to include additional deployment options in the future.

Main Capabilities

Complex GUI Automation: Autonomously complete complex interface operations containing hundreds of interactive elements
Cross-System Data Integration: Extract and integrate multi-source data through pure visual interaction without API interfaces
Long-Task Planning Execution: Support enterprise-level business process automation of dozens to hundreds of steps
Intelligent Report Generation: Automatically generate structured documents such as data analysis reports and work summaries
Edge-Native Inference: Efficient on-device execution on Apple Silicon via INT8 activation quantization (Cider)
Autonomous Application Construction: Drives end-to-end software construction pipelines through visual GUI operation (Mano-AFK)

Technical Background

Mano-P builds upon the complete technical framework of the Mano project (see Mano Technical Report), employing the Mano-Action bidirectional self-reinforcement learning method, three-stage progressive training (SFT → Offline Reinforcement Learning → Online Reinforcement Learning), "think-act-verify" loop reasoning mechanism, and a closed-loop data circulation system to achieve high-precision GUI understanding and operation capabilities. The edge version is optimized through mixed-precision quantization, visual token pruning, and edge inference adaptation, enabling large-scale parameter models to run efficiently on edge devices like Mac mini/MacBook/computing sticks.

🎯 Key Highlights

#1 on OSWorld Benchmark: Mano-P 1.0-72B achieves 58.2% success rate on OSWorld, ranking first among all specialized GUI agent models, outperforming the second-place opencua-72b (45.0%) by 13.2 percentage points
Leading on WebRetriever Protocol I: Mano-P 1.0 scores 41.7 NavEval, surpassing Gemini 2.5 Pro Computer Use (40.9) and Claude 4.5 Computer Use (31.3)
Fully Local Execution: Runs inference locally on Apple M4 chip with 32GB RAM (Mac mini or MacBook). No cloud API calls required. All screenshots and task data stay on-device
High-Performance Inference: Mano-P 1.0-4B achieves ~80 tokens/s decode on Apple M5 Pro; with Cider's W8A8 activation quantization, prefill speeds up by ~12.7% over the W8A16 baseline
Autonomous Long-Task Execution: Supports complex business processes with end-to-end automation without internet connectivity
Edge-Native INT8 Acceleration: Companion Cider SDK adds the W8A8 / W4A8 activation-quantization primitives MLX lacks natively, delivering 1.4x–2.2x prefill speedup over MLX W4A16 on Apple M5 Pro — works with any MLX model, not just Mano-P
Autonomous Software Construction: Mano-AFK drives a full PRD → code → deploy → test → fix loop using Mano-P as its local vision model for real-browser E2E testing — from a single natural-language prompt to a deployed, tested application, no human in the loop

🎬 Use Case Demonstrations

Scenario 1: Mano-AFK Fully automated application construction

Mano-P_AFK_EN.mp4

We demonstrated the fully automated application construction process of Mano-AFK. After receiving natural language requirements, the system sequentially completes requirement clarification, technical architecture design, code generation, local deployment, and multi-level testing (API interface testing, LLM based page visual inspection, and end-to-end GUI automation testing driven by VLA model). When the test fails, the system automatically locates the root cause of the problem, fixes the code, and deploys verification again, iterating until all test cases pass. The entire process does not require manual intervention, and ultimately delivers a runnable application with complete requirement documents and build reports.

Scenario 2: Commercial video intelligent system

Mano-P.Commercial_video_intelligent_system_EN.mp4

We fully demonstrated the actual workflow of a commercial video intelligent system. Starting from the user's command, the system automatically completes the entire process of video generation, uploading, analysis, editing, and secondary evaluation. During the process, the system can autonomously operate web pages and editing software, complete fine operations such as file processing and subtitle modification, and generate analysis reports containing subjective evaluations and objective indicators. By comparing the differences between the initial and refined versions, visually present the overall capabilities and application effects of the system.

Scenario 3: Local model task execution

Mano-P._EN.mp4

Mano-P, The small-sized end side GUI-VLA model can run directly on your computer, supporting direct inference operation on Macmini/Macbook with M4 chip and above, as well as direct operation on plug and play computing power sticks. In the CUA scenario, break through the bottleneck of human participation in the Agent workflow. Mano-P, The first step in leading Private AI.

Scenario 4: Daily Life and Entertainment Applications

Mano_._EN.mp4

Mano-P excels not only in enterprise-level business automation but also integrates seamlessly into daily life. This video demonstrates the system's application in Mahjong gameplay: through pure visual understanding of the game interface, it autonomously completes tile recognition, analysis, and decision-making. This case validates Mano-P's general-purpose capabilities beyond work scenarios—from office automation to leisure entertainment, from structured data processing to unstructured game interactions, truly realizing the vision of "Private AI." One model, adapting to every aspect of life and work.

Scenario 5: Mano-AFK × Cider Locally-Accelerated End-to-End App Construction

0429_Mano-AFK+Cider_EN_compressed.mp4

This video demonstrates the combined capability of Mano-AFK and the Cider inference acceleration SDK. Starting from a single natural-language requirement, Mano-AFK autonomously performs requirement clarification, architecture design, code generation, and local deployment. In the E2E testing stage it drives a real browser through the local Mano-P vision model — accelerated by Cider — to run GUI automation tests; on failure, it localizes the defect, patches the code, and re-verifies until a runnable application is delivered. Cider supplies the INT8 activation-quantization primitives that give Mano-P a substantial prefill speedup on Apple Silicon, so the entire build–test–fix loop runs fully on-device, balancing autonomy, privacy, and performance.

📊 Benchmark Performance

Performance of the Mano series models in multiple benchmarks:

1. GUI Grounding

📊 Expand Evaluation Data

2. BUA & CUA

📊 Expand Evaluation Data

OSWorld - Specialized Models

OSWorld - All Models

WebRetriever

3. Perception & Cognition

📊 Expand Evaluation Data

Video-SME-2

Models	Protocol	CA		CV		PAR		Saliency
Models	Protocol	Acc	F1	Acc	F1	Acc	F1	KL↓	CC↑	SIM↑	NSS↑	AUC↑
Random	P1	10.42	11.03	10.76	10.95	15.94	16.00	2.1789	0.0452	0.2852	0.1081	0.5340
Random	P2	10.01	10.74	10.32	10.50	14.39	15.04	4.3378	0.0270	0.2274	0.0665	0.5273
Zero-shot for MLLMs
GPT4o	P1	15.17	6.57	16.11	9.58	16.71	10.34	1.9423	0.4660	0.4602	1.2842	0.7848
GPT4o	P2	10.26	4.77	12.16	7.66	15.00	8.55	2.2650	0.4097	0.4028	1.2418	0.7807
Gemini 2.0 Flash	P1	17.18	5.13	25.06	8.39	24.94	9.52	1.4726	0.3380	0.3751	0.8629	0.7296
Gemini 2.0 Flash	P2	10.45	4.26	12.60	4.95	15.96	7.90	1.6373	0.3542	0.3490	1.0027	0.7590
GPT-5.2	P1	17.83	7.67	22.22	12.55	16.17	9.74	1.3262	0.4852	0.4632	1.3078	0.7969
GPT-5.2	P2	15.31	5.14	19.88	10.27	13.56	7.42	1.5444	0.4379	0.4092	1.3006	0.7999
Claude Sonnet 4.5	P1	10.34	5.8	13.26	9.84	16.02	9.94	1.4235	0.4912	0.4213	1.2956	0.8042
Claude Sonnet 4.5	P2	10.34	5.55	13.27	7.08	16.02	9.6	1.2855	0.4564	0.4781	1.3112	0.7915
Llama 4 Scout	P1	13.98	9.96	10.25	6.51	13.27	8.11	3.7166	0.3331	0.3849	0.8828	0.7238
Llama 4 Scout	P2	10.00	7.33	11.10	8.49	14.35	7.42	3.7434	0.3019	0.3452	0.8848	0.7258
Qwen2.5-VL-7B	P1	15.88	5.21	10.07	6.07	12.26	4.96	12.0586	0.0999	0.2154	0.2578	0.5852
Qwen2.5-VL-7B	P2	10.25	3.95	10.89	5.83	14.39	5.73	12.7596	0.0762	0.1855	0.2195	0.5753
InternVL3-8B	P1	13.35	7.78	14.71	8.02	10.20	6.95	12.6480	0.0572	0.1895	0.1140	0.5769
InternVL3-8B	P2	10.58	6.70	10.94	8.12	12.68	6.32	12.1385	0.0604	0.1819	0.1395	0.5859
Fine-tune for MLLMs
Qwen2.5-VL-7B	P1	22.51	19.11	23.39	10.83	32.06	25.88	1.5091	0.6953	0.6118	1.8937	0.8579
Qwen2.5-VL-7B	P2	13.72	13.25	13.03	10.94	21.24	20.65	2.2496	0.5359	0.4793	1.6439	0.8221
InternVL3-8B	P1	20.94	18.41	21.96	11.02	30.33	24.66	1.2551	0.7014	0.6340	1.9896	0.8670
InternVL3-8B	P2	12.81	11.83	12.16	11.11	19.26	19.27	1.8759	0.6282	0.5467	2.0621	0.8627
Mano-P 1.0
Stage I	P1	31.27	30.53	27.31	25.18	35.16	34.45	0.6794	0.7670	0.7015	2.1347	0.8710
Stage I	P2	21.89	22.06	18.27	18.57	23.77	23.87	1.5759	0.6482	0.6167	2.1021	0.8627
Stage II	P1	32.59	31.46	27.57	25.76	37.73	35.79	0.6736	0.7686	0.7120	2.1688	0.8853
Stage II	P2	20.55	21.26	15.37	15.15	25.36	25.83	0.5617	0.6440	0.6130	2.1090	0.8602
Stage III	P1	34.58	33.99	31.92	28.37	39.42	37.63	0.6073	0.7853	0.7248	2.2103	0.8938
Stage III	P2	25.29	25.83	20.21	19.29	26.49	26.54	1.4617	0.6725	0.6330	2.1788	0.8776

MIT1003 & SalECI

Dataset	Method	Saliency
Dataset	Method	KL↓	CC↑	SIM↑	NSS↑	AUC↑
MIT1003	FastSal	1.036	0.590	0.478	2.008	0.875
	SAM-Resnet	1.247	0.746	0.597	2.752	0.902
	DAV	0.753	0.699	0.566	2.574	0.897
	UNISAL	1.014	0.734	0.597	2.759	0.902
	Transalnet	0.660	0.722	0.592	2.631	0.903
	SUM	0.563	0.768	0.630	2.839	0.913
	Mano-P 1.0	0.648	0.770	0.698	2.950	0.902
SalECI	SSM	0.720	0.599	0.611	1.396	0.830
	DeepGaze IIE	0.995	0.560	0.399	1.327	0.842
	EML-NET	1.220	0.510	0.536	1.232	0.807
	Transalnet	0.873	0.717	0.534	1.723	0.824
	Temp-Sal	0.712	0.719	0.629	1.768	0.813
	SSwinTransformer	0.652	0.687	0.606	1.701	0.868
	Mano-P 1.0	0.615	0.769	0.695	1.735	0.868

ETMD

Saliency Metrics

Methods	Saliency
Methods	CC ↑	SIM ↑	NSS ↑	AUC ↑
ACLNet	0.477	0.329	2.36	0.915
TASED-Net	0.479	0.366	2.63	0.916
STAViS	0.569	0.425	2.94	0.931
ViNet	0.569	0.409	3.06	0.928
CASP-Net	0.620	0.478	3.34	0.940
Mano-P 1.0	0.642	0.481	2.99	0.929

Emotion Recognition

	Emotion Valence		Emotion Arousal
	Acc ↑	Acc ± 1 ↑	Acc ↑	Acc ± 1 ↑
Qwen2.5-VL-7B	13.3	38.1	10.8	35.5
Mano-P 1.0	20.2	46.5	18.7	47.3

4. Pruning

📊 Expand Evaluation Data

Online-Mind2Web

Comparison of Task Execution Success Rate (SR) on Online-Mind2Web Benchmark Avg. Tokens/img represents the average visual token retention rate per image; lower values indicate more aggressive pruning.

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.

Model	Method	Avg. Tokens /img ↓	Training samples/s ↑	SR (↑)
Qwen3VL-2B	Baseline (w/o FT)	100%	5.08	0.290
	Baseline (FT)	100%	5.09	0.390
	TextGuide	12.55%	13.54	0.310
	FlashVLM [4]	12.55%	17.01	0.343
	Compressor-VLA [11]	13.33%	16.92	0.293
	HiPrune [16]	25.09%	16.67	0.333
	PDrop [33]	41.47%	10.43	0.330
	IVC	25.09%	7.89	0.303
	Mano-P 1.0	25.09%	20.04	0.370
	Mano-P 1.0	12.57%	22.62	0.336
Qwen3VL-4B	Baseline (FT)	100%	3.24	0.425
	PDrop	41.47%	5.58	0.365
	IVC	25.09%	4.67	0.343
	GSPruning	25.09%	16.72	0.400

5. Context Learning

📊 Expand Evaluation Data

CL Bench

🔧 Skills

Mano-Skill is a desktop GUI automation tool based on the Mano model, driving cross-platform graphical interface operations through natural language. We provide two different usage forms for the same core capability to adapt to different usage scenarios and user groups.

📦 Core Capability Overview

Feature Highlights

Natural Language Driven: Users describe tasks in natural language, and the system automatically executes GUI operations
Flexible Inference Modes:
- Local Mode: Models run locally, data stays on device, fast response
  - Run directly on Mac mini/MacBook (M4 chip or above, 32GB+ RAM)
  - Or use Mano-P computing stick (via USB 4.0 connection)
- Cloud Mode: Without local model configuration, uses cloud API service (mano.mininglamp.com)
- System automatically detects local model configuration and seamlessly switches inference modes
Comprehensive Interaction Support: Click, type, hotkey, scroll, drag, mouse movement, screenshot, wait, app launch, URL navigation
Cross-Platform Support: macOS (stable), Windows, Linux (Beta)

How It Works

Cloud Mode (Default)

Capture current screen screenshot
Send screenshot and task description to cloud vision model (mano.mininglamp.com)
Cloud model analyzes and returns next action instruction
Local client executes operation (click, type, etc.)
Loop execution until task completion

Local Mode

Capture current screen screenshot
Run Mano-P model on local device (Mac mini/MacBook) or computing stick for inference
Local model analyzes and returns next action instruction
Client executes operation (click, type, etc.)
Loop execution until task completion

Data Privacy & Security

Cloud Mode:

⚠️ Data Sent: Screenshots and task descriptions sent to mano.mininglamp.com for real-time visual analysis
✅ Data Not Sent: Does not access or transmit local files, clipboard contents, system credentials
⚠️ Privacy Note: Avoid displaying sensitive documents, chat logs, or credential information on screen when running tasks

Local Mode (Mac mini/MacBook or Computing Stick):

✅ Fully Local Processing: All data processing is completed locally, screenshots and task descriptions never leave the device
✅ Data Stays on Device: Does not access or transmit any data to external servers
✅ Maximum Privacy Protection: Suitable for handling sensitive information and high-security scenarios

General Assurance:

✅ Open Source Auditable: Complete source code publicly available for review

🔧 Two Usage Forms

If you want to use Mano-P directly to accomplish GUI automation tasks, here are two different usage forms. Choose the one that best fits your use case.

1️⃣ mano-cua (CLI Command-Line Tool)

Use Case: For human users — developers and advanced users invoking mano-cua directly in a terminal to quickly execute one-off or scripted GUI automation tasks

Installation:

# Install via Homebrew
brew tap Mininglamp-AI/tap
brew install mano-cua

The installation process will automatically:

Create an isolated Python 3.13 virtual environment
Install required dependencies (including Tkinter GUI library)
Configure the executable command to system PATH

Usage:

# Cloud mode (default — no extra setup required)
mano-cua run "Open WeChat and tell FTY the meeting is postponed"
mano-cua run "Search for AI news on Xiaohongshu and display the first post"

# Stop current task
mano-cua stop

mano-cua ships with two inference modes: cloud and local. Cloud is the default; to run Mano-P locally on macOS Apple Silicon, use the --local flag:

# First-time local setup: verify env / install SDK / pull the local model
mano-cua check
mano-cua install-sdk
mano-cua install-model

# Run a task in local mode
mano-cua run "Open Safari and search for Python" --local
mano-cua run "Type hello in the search box" --local --url "https://www.baidu.com" --minimize --max-steps 15

In local mode, Mano-P runs on-device via MLX — screenshots and task descriptions make zero network calls for inference.

Features:

✅ Command-line interface, quick invocation
✅ Virtual environment isolation, no system Python pollution
✅ Suitable for script integration and batch processing
✅ Can be embedded in shell scripts
✅ Cloud / local inference modes — --local switches to on-device in one flag

Installation & Distribution:

Homebrew Tap: github.com/Mininglamp-AI/homebrew-tap

2️⃣ mano-skill (ClawHub Skill Form)

Use Case: For AI agents — Claude Code, OpenClaw and similar agents autonomously invoke GUI automation capabilities mid-reasoning to complete user tasks, no manual command execution needed

Installation:

Option 1: Install via Claude Code

In Claude Code, skills exist as "commands". Installation steps:

Download the skill zip package from ClawHub
After extraction, copy files to Claude Code's commands directory
Restart Claude Code or in a new session, the skill will be automatically available

Option 2: Install via ClawHub CLI (Recommended)

Use the ClawHub CLI tool for one-click installation and skill management:

# Install skill
clawhub install mano-cua

# Install specific version
clawhub install mano-cua --version 1.0.0

# Update skill to latest version
clawhub update mano-cua

After installation, start a new Claude Code or OpenClaw session to use.

Prerequisites: ClawHub CLI tool must be installed first. See: OpenClaw Documentation - ClawHub

Usage:

When users make requests to AI agents that require GUI operations, the agent will automatically invoke this skill:

User: "Help me open WeChat, find FTY's chat window, and tell him the meeting is postponed to tomorrow"
Agent: [Automatically invokes mano-skill to complete GUI operation]

Features:

✅ Autonomously invoked by AI agents, no manual command execution needed
✅ Deeply integrated with agent reasoning capabilities
✅ Suitable for complex multi-step task automation
✅ ClawHub ecosystem with version management and security scanning

Project Resources:

Source Code: github.com/Mininglamp-AI/mano-skill
ClawHub Home: clawhub.ai/hanningwang/mano-cua
Version: v1.0.0
License: MIT

⚙️ Permission Requirements (Common to All Forms)

Screen Recording Permission
Accessibility Permission (keyboard/mouse control)
Grant permissions in System Preferences → Privacy & Security

🔒 Security Constraints (Common to All Forms)

Sensitive or potentially dangerous operations require user confirmation before execution
Users can stop tasks at any time
Only one task can run on each device simultaneously
Only supports primary display (multi-display environment)

📊 Status Panel

When a task is running, a small status panel appears in the top-right corner of the screen to:

Display real-time task status and progress
Provide task management functions (pause/stop)
Remind users that an automation task is running to avoid accidental interference

🔔 Platform Compatibility Note

Beta Version Notice: Mano-Skill is currently in Beta testing phase.

macOS: ✅ Preferred and most thoroughly tested platform, stable and ready for use
Windows and Linux: ⚠️ Platform adaptations not yet fully completed, minor issues may occur

We are continuously improving cross-platform compatibility. Feedback is welcome.

🤖 Models

If you want to integrate Mano-P's model capabilities into your own applications, this section provides performance metrics and usage guidelines.

Performance Evaluation

The table below presents actual inference benchmark results of Mano-P 1.0-4B running on Apple M5 Pro with the Cider inference SDK. Using W8A16 (MLX's native weight-only quantization path) as the baseline — the same reference convention adopted in Cider's quantization benchmark — enabling Cider's W8A8 activation quantization reduces prefill time from 2.839s to 2.519s on the same input, a ~12.7% prefill speedup. For more data, refer to the ⚡ Inference SDK section below.

Model	Chip	Bandwidth	Framework	Context Length	Quantization	Prefill Time (s)	Decode Speed (tokens/s)
Mano-P 1.0-4B	Apple M5 Pro 64GB RAM	307 GB/s	Cider	4516	W8A16	2.839	80.1
Mano-P 1.0-4B	Apple M5 Pro 64GB RAM	307 GB/s	Cider	4516	W8A8	2.519	79.5

Model download: 🤗 Hugging Face · 🪄 ModelScope

⚡ Inference SDK

Overview

Cider is an inference acceleration SDK developed on top of MLX for macOS. It provides online activation quantization operators absent in MLX, with custom int-matmul kernels built as MLX custom primitives supporting full lazy evaluation. It also includes service-side extensions and non-intrusive compatibility patches for mlx_vlm (validated on mlx_vlm 0.4.3), including fixes for Qwen3-VL multi-image inference issues related to RoPE position handling and chunked prefill.

Conditional Compilation (M4 / M5)

Cider uses conditional compilation: the INT8 TensorOps C++ extension is only built on Apple M5+.

Chip	`pip install -e .` behavior	`import cider` behavior
M5+	Full build (CMake + Metal kernels)	All features available
M4 and below	Skips C++ build, installs pure-Python package	`is_available()` → False, `convert_model()` is a warning no-op

Override via environment variable:

CIDER_FORCE_BUILD=1 pip install -e .   # Force build (e.g., CI)
CIDER_FORCE_BUILD=0 pip install -e .   # Force skip

Modes

Mode	Weights	Activations	Compute Path	Status
W8A8	INT8 symmetric	INT8 per-token	TensorOps matmul2d	✅ Implemented
W4A8	INT4 packed (uint8)	INT8 per-token	Unpack → TensorOps	✅ Implemented
W4A16	—	—	MLX built-in	Baseline
W8A16	—	—	MLX built-in	Baseline

W4A16 and W8A16 are already supported by MLX natively — this SDK provides the missing W8A8 and W4A8 modes that MLX does not implement.

MLX's quantization is weight-only: QuantizedLinear dequantizes weights to FP16 and uses FP16 GEMM. While MLX's Steel NAX templates are generic enough to be instantiated with INT8 types (and would achieve identical raw matmul throughput — see the transparent benchmark), MLX does not provide the quantization/dequantization pipeline needed for actual W8A8 inference. Cider fills this gap with fused quantize-matmul-dequant primitives, implementing online INT8 activation quantization and INT8 TensorOps-based compute for the supported inference paths.

W8A8 Quantization Granularity

Granularity	Description	Speed	Precision
Per-channel	One scale per output channel	Fastest (1.8x prefill)	Slightly lower
Per-group (gs=128)	One scale per 128 elements	Fast (1.5x prefill)	Moderate precision retention
Per-group (gs=64)	One scale per 64 elements	Moderate (1.3x prefill)	Higher precision

Performance (Apple M5 Pro)

Individual Operator Latency

Shape [N=10240, K=2560]

M	PC(ms)	PG(ms)	w8a16	w4a16	PC/w8	PC/w4	PG/w8	PG/w4
1	0.27ms	0.26ms	0.26ms	0.18ms	0.96x	0.67x	0.99x	0.69x
128	0.34ms	0.39ms	0.49ms	0.44ms	1.43x	1.28x	1.26x	1.13x
1024	1.23ms	1.52ms	2.24ms	2.04ms	1.82x	1.66x	1.47x	1.34x
4096	4.41ms	5.65ms	8.12ms	7.72ms	1.84x	1.75x	1.44x	1.37x
8192	8.71ms	11.40ms	16.23ms	15.09ms	1.86x	1.73x	1.42x	1.32x

Shape [N=2560, K=10240]

M	PC(ms)	PG(ms)	w8a16	w4a16	PC/w8	PC/w4	PG/w8	PG/w4
1	0.25ms	0.26ms	0.26ms	0.20ms	1.03x	0.78x	0.98x	0.75x
128	0.39ms	0.41ms	0.55ms	0.46ms	1.43x	1.19x	1.35x	1.12x
1024	1.31ms	1.65ms	2.35ms	2.14ms	1.80x	1.64x	1.43x	1.30x
4096	5.37ms	6.79ms	8.54ms	8.04ms	1.59x	1.50x	1.26x	1.18x
8192	10.97ms	12.94ms	17.28ms	16.23ms	1.58x	1.48x	1.34x	1.25x

End-to-End VLM

Qwen3-VL-2B

Prompt Tokens	FP16 Prefill (tok/s)	W8A16 Prefill (tok/s)	W8A8 PC Prefill (tok/s)	FP16 Decode (tok/s)	W8A16 Decode (tok/s)	W8A8 PC Decode (tok/s)
1334	3010	2065	3242	70	107	104
2393	2868	1847	2983	69	97	100
3455	2777	1741	2796	66	90	95

Qwen3-VL-4B

Prompt Tokens	FP16 Prefill (tok/s)	W8A16 Prefill (tok/s)	W8A8 PC Prefill (tok/s)	FP16 Decode (tok/s)	W8A16 Decode (tok/s)	W8A8 PC Decode (tok/s)
1334	1884	1786	2186	32	56	54
2393	1815	1700	2028	31	55	52
3455	1755	1603	1881	30	52	49

LLM Quantization: Precision vs. Speed Comparison

Models	Quantization Configuration	wikitext2 PPL (↓)	Prefill Time (s) (↓)	Peak Memory (GB) (↓)
Qwen3-8B	FP16	9.726	179.9	18.93
	W8A16 (mlx RTN)	9.707	221.3	12.07
	W8A8 (per-channel)	9.756	123.5	11.32
	W8A8 (per-group gs=64)	9.744	179.1	11.83
	W8A8 (per-group gs=128)	9.727	165.8	11.61
Llama3-8B	FP16	6.138	175.8	18.32
	W8A16 (mlx RTN)	6.147	236.9	11.46
	W8A8 (per-channel)	6.271	123.3	10.69
	W8A8 (per-group, gs=64)	6.269	178.7	11.19
	W8A8 (per-group, gs=128)	6.270	155.7	10.98

🔗 Repository: github.com/Mininglamp-AI/cider

⚗️ Approach

If you are a researcher or wish to train customized GUI Agent models based on your own data, we plan to open-source the complete Mano-Action training methodology and related tools.

Release Soon

Mano-Action Training Methodology

Mano-Action is a bidirectional self-reinforcement training framework specifically designed for GUI Grounding. Unlike traditional unidirectional prediction methods, Mano-Action achieves more robust interface understanding through Text↔Action cycle consistency learning, enabling the model to master both "locating elements from descriptions" and "describing given elements" simultaneously.

Core Features

Bidirectional Cycle Learning: Mutual reinforcement between Text → Action and Action → Text
Three-Stage Progressive Training: Supervised Learning → Offline RL → Online RL
Closed-Loop Data Generation: Automatically generate high-quality training data for continuous model improvement
Edge Optimization Adaptation: Includes quantization, pruning, and other edge deployment optimization techniques

Use Cases

🎓 Academic Research: Explore new approaches to GUI understanding and multimodal interaction
🏢 Enterprise Customization: Train specialized models based on internal enterprise systems
🌐 Domain Adaptation: Fine-tune models for specific domains (healthcare, finance, etc.)
🔬 Algorithm Innovation: Develop new training techniques building on Mano-Action

🌟 Technical Advantages

Mano-P vs Traditional Solutions CUA Comparison

Feature	Mano-P	OpenClaw	Manus	Traditional RPA
Model Source	✅ Built-in edge model	⚠️ User-configured	⚠️ Cloud API calls	❌ No model (rule-based engine)
Data Security	✅ Local execution	⚠️ LLM/skill cloud calls	⚠️ Cloud inference	✅ Can be local
Control Method	✅ Pure visual	⚠️ CDP protocol+CLI	❌ HTML parsing+CLI	❌ System API
Use Scenarios	✅ All-type GUI (desktop/Web/3D)	✅ Multi-type apps	⚠️ Web apps only	⚠️ Specific systems
Long Task Plan	✅ Autonomous planning	✅ Autonomous planning	✅ Visual flow orchestration	❌ Needs preset workflows
Response Speed	✅ Instant response	✅ Local/cloud execution	⚠️ Cloud latency	✅ Instant response
Deployment Cost	✅ Low-cost entry	✅ Open source & free	⚠️ Subscription fee	✅ Low cost
Robustness	✅ UI change adaptive	✅ LLM adaptive	⚠️ Limited adaptation	❌ UI change needs reconfig

Core Competitiveness

Edge Large Model + Flexible Deployment
- 4B model runs directly on Mac (M4 chip + 32GB RAM)
- Large parameter models (72B) supported via computing stick
- No API key configuration needed, ready out-of-the-box
- Significant advantage over OpenClaw (requires user model configuration) and Manus (cloud calls)
Universal Visual Understanding
- Pure visual GUI interaction, not limited to browsers and web apps
- Broader support than OpenClaw (CDP protocol mainly for browsers) and Manus (web apps only)
- Supports desktop software, 3D applications, professional tools, and non-standard GUIs
Offline Long-Task Autonomous Planning
- Fully offline reasoning for complex business processes
- Autonomous decision-making and error correction without internet connection
- Unique advantage over Manus (cloud latency) and traditional RPA (needs preset workflows)
Integrated Hardware Deployment
- Model + computing stick integrated solution, plug-and-play
- Lowers technical barrier compared to OpenClaw (open-source & free but requires self-deployment)
- Multiple deployment forms (direct Mac install / plug-and-play compute stick), rapid launch

🔗 Applications

Mano-AFK — an autonomous full-cycle app builder that turns a single natural-language sentence into a deployed, tested, and bug-fixed application. Its E2E testing stage runs by default with Mano-P as the local backend (screenshots and task descriptions stay on-device); it can also switch to cloud mode driven by Anthropic's Claude CUA (requires ANTHROPIC_API_KEY; third-party service with no local alternative). Mano-AFK is a concrete application scenario of Mano-P in real-world software engineering pipelines.

GitHub: github.com/Mininglamp-AI/mano-afk
ClawHub: clawhub.ai/hanningwang/mano-afk

CUA Benchmark — Mano-P 4B on MacBook Pro M5 (16GB)

The suite evaluates 100 tasks across 5 web applications that were themselves built autonomously by Mano-AFK: TripSplit (expense splitting), md-wechat (Markdown → WeChat formatter), OMS (order management), Family Ledger (household bookkeeping), and Life Dashboard (personal widgets). Each app ships in two variants — a golden build (bug-free, expected verdict PASS, 76 tasks) and a buggy build with specific UI/logic defects injected (expected verdict FAIL, 24 tasks). Accuracy is defined as the share of tasks where the judge's verdict matches the expected label; each project contributes 15–16 golden tasks and 4–5 bug-injection tasks.

Configuration	Accuracy	Avg Steps	Prefill Speed	Avg Tokens/Step
W8A16	58.0%	6.1	~1,253 tok/s	3,389
W8A8 (Cider)	54.0%	6.93	~1,453 tok/s	3,104

Metrics: Accuracy — judge verdict matches expected PASS/FAIL; Steps — actions taken per task; Prefill Speed — token throughput during the prefill phase; Tokens/Step — prompt + generation tokens combined.

Note on W8A8 on this hardware: The W8A8 row runs via Cider for INT8 activation quantization. W8A8 accelerates prefill through INT8 TensorOps, but it must hold both the original and the INT8 weights in memory simultaneously — roughly doubling weight footprint. On a 16 GB unified-memory device, the added pressure can trigger swapping that offsets the prefill gain, which is why W8A8 here lands slightly behind W8A16 on accuracy. At least 4 GB of free memory beyond the model size is recommended to see W8A8's full benefit.

Benchmark methodology — full test design, task schema, and bug catalog

📄 Technical Papers & Citation

Related Papers

Mano-P is based on the following research work:

1. Mano Series Model Foundation Paper

@article{mano-2025,
  title={Mano Technical Report},
  author={Tianyu Fu, Anyang Su, Chenxu Zhao, Hanning Wang, Minghui Wu, Zhe Yu, Fei Hu, Mingjia Shi, Wei Dong, Jiayao Wang, Yuyang Chen, Ruiyang Yu, Siran Peng, Menglin Li, Nan Huang, Haitian Wei, Jiawei Yu, Yi Xin, Xilin Zhao, Kai Gu, Ping Jiang, Sifan Zhou, Shuo Wang},
  journal={arXiv preprint arXiv:2509.17336},
  year={2025},
  url={https://arxiv.org/abs/2509.17336}
}

2. WebRetriever Benchmark

@article{webretriever-2026,
  title={WebRetriever: A Large-Scale Comprehensive Benchmark for Efficient Web Agent Evaluation},
  author={Wei Dong and Tianyu Fu and Zhe Yu and Hanning Wang and Anyang Su and Zhizhou Fang and Yuyang Chen and Shuo Wang and Minghui Wu and Ping Jiang and Zhen Lei and Chenxu Zhao},
  year={2026},
  note={To be published},
  url={https://github.com/hhhhhhalf/WebRetriever}
}

Academic Collaboration

We welcome collaboration with academia:

🔬 Dataset Contribution: Provide new GUI task datasets
🤝 Joint Research: Collaborate on edge deployment, quantization optimization, GUI understanding, etc.
📚 Benchmarking: Test Mano-P on new evaluation sets

For academic collaboration inquiries, please contact: model@mininglamp.com

❓ FAQ

🤖 What is Mano-P?

Mano-P is an open-source GUI-VLA (Vision-Language-Action) agent designed to run locally on Apple Silicon edge devices. It uses pure visual understanding to automate desktop GUI operations across platforms.

⚖️ How does Mano-P compare to Claude Computer Use?

Performance Comparison:

OSWorld (all models): Claude Sonnet 4.6 72.1% vs Mano-P 1.0-72B 58.2%
WebRetriever Protocol I: Mano-P 41.7 NavEval vs Claude 4.5 Computer Use 31.3

Key Difference:

✅ Mano-P runs entirely on-device, no data leaves the machine
⚠️ Claude Computer Use requires cloud API calls

Use Case: Mano-P is particularly suitable for high-security scenarios.

🔌 Can Mano-P run without internet?

Yes! In local mode, all model inference runs on the Apple M4 device. ✅ No screenshots or task descriptions are sent to external servers.

💻 What hardware do I need?

Minimum Requirements:

Mac mini or MacBook
Apple M4 chip
32GB RAM

Alternative:

Any Mac + Mano-P computing stick (connected via USB 4.0+)

📌 We plan to support more devices in the future.

📦 How do I install Mano-P?

CLI Tool:

brew tap Mininglamp-AI/tap && brew install mano-cua

OpenClaw/Claude Code Skill: See ClawHub - Mano-CUA

🔒 Is my data safe?

Local Mode: ✅ All processing happens on-device

Cloud Mode:

⚠️ Only screenshots and task descriptions sent to mano.mininglamp.com
✅ No local files, clipboard contents, or credentials accessed

Transparency: Full client is open-source for audit

🤝 Contribution Guidelines

We welcome community contributions! If you want to contribute to the project:

Fork this repository
Create your feature branch (git checkout -b feature/AmazingFeature)
Commit your changes (git commit -m 'Add some AmazingFeature')
Push to the branch (git push origin feature/AmazingFeature)
Open a Pull Request

Contribution Areas

🐛 Bug fixes and issue reporting
📝 Documentation improvements and translations
💡 New feature suggestions and implementations
🧪 Test cases and benchmarking
🎨 Application scenarios and demo contributions

📄 License

This project is licensed under the Apache License 2.0.

License Highlights:

✅ Commercial use
✅ Modification and distribution
✅ Patent grant
⚠️ Must retain copyright notice
⚠️ Must state changes

📮 Contact

📧 Email: model@mininglamp.com
🏠 Website: https://github.com/Mininglamp-AI/Mano-P
💬 Community: (To be added)
🐛 GitHub Issues: https://github.com/Mininglamp-AI/Mano-P/issues

🙏 Acknowledgments

Thanks to all developers and researchers who contributed to this project.

Special Thanks:

Mano project team for providing the technical foundation
DeepMiner platform for deep integration support
Edge computing hardware partners
Open source community contributors

_{Built with ❤️ by the Mano-P Team}

Name		Name	Last commit message	Last commit date
Latest commit History 47 Commits
pics		pics
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
README_CN.md		README_CN.md

Folders and files

Latest commit

History

Repository files navigation

Mano-P 1.0

📖 Overview

Main Capabilities

Technical Background

🎯 Key Highlights

🎬 Use Case Demonstrations

Scenario 1: Mano-AFK Fully automated application construction

Scenario 2: Commercial video intelligent system

Scenario 3: Local model task execution

Scenario 4: Daily Life and Entertainment Applications

Scenario 5: Mano-AFK × Cider Locally-Accelerated End-to-End App Construction

📊 Benchmark Performance

1. GUI Grounding

2. BUA & CUA

OSWorld - Specialized Models

OSWorld - All Models

WebRetriever

3. Perception & Cognition

Video-SME-2

MIT1003 & SalECI

ETMD

Saliency Metrics

Emotion Recognition

4. Pruning

Online-Mind2Web

5. Context Learning

CL Bench

🔧 Skills

📦 Core Capability Overview

Feature Highlights

How It Works

Data Privacy & Security

🔧 Two Usage Forms

1️⃣ mano-cua (CLI Command-Line Tool)

2️⃣ mano-skill (ClawHub Skill Form)

⚙️ Permission Requirements (Common to All Forms)

🔒 Security Constraints (Common to All Forms)

📊 Status Panel

🔔 Platform Compatibility Note

🤖 Models

Performance Evaluation

⚡ Inference SDK

Overview

Conditional Compilation (M4 / M5)

Modes

W8A8 Quantization Granularity

Performance (Apple M5 Pro)

⚗️ Approach

Mano-Action Training Methodology

Core Features

Use Cases

🌟 Technical Advantages

Mano-P vs Traditional Solutions CUA Comparison

Core Competitiveness

🔗 Applications

📄 Technical Papers & Citation

Related Papers

Academic Collaboration

❓ FAQ

🤝 Contribution Guidelines

Contribution Areas

📄 License

📮 Contact

🙏 Acknowledgments

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Packages