Observed Signal · Apr 24, 2026 · Technical Release · Source: DEV Community · Impact: 2/5 · Sentiment: Positive
Mano-P: Edge-Native AI Agent Restores Data Sovereignty
The article presents Mano-P, an open-source, edge-native AI agent architecture designed to run entirely on local hardware to preserve data sovereignty and reduce cloud dependencies. Mano-P uses vision-only understanding (screenshots as raw pixels), w4a16 quantization, and GS-Pruning to run a 4B-parameter model interactively on consumer Apple Silicon. Measured on an Apple M4 Pro (32GB), the model shows 476 tokens/s prefill, 76 tokens/s decode, and 4.3 GB peak memory. Benchmarks cited include a 58.2% success rate on OSWorld and 41.7 NavEval on WebRetriever Protocol I, outperforming larger cloud models in GUI automation tasks. The project follows a three-stage training pipeline (SFT, offline RL, online RL), supports local USB 4.0 accelerator offload, and is being released in phased open-source stages under Apache 2.0.
Demonstrates practical, open-source edge inference for AI agents that preserves data sovereignty and shows competitive benchmarks on-device; relevant to privacy-sensitive applications and on-device inference trends but not a major-platform or market-moving announcement.
Track Apple Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- Mano-P is an open-source GUI-aware agent released under the Apache 2.0 license (project on GitHub: Mininglamp-AI/Mano-P).
- Mano-P is a 4B-parameter model using w4a16 quantization; on an Apple M4 Pro (32GB) it measured 476 tokens/s prefill, 76 tokens/s decode, and 4.3 GB peak memory.
- Benchmarks: Mano-P achieved a 58.2% success rate on OSWorld and 41.7 NavEval on WebRetriever Protocol I, outperforming listed larger cloud models in GUI automation.
- Architecture and training: vision-only input (screenshots), three-stage training (Supervised Fine-Tuning, Offline RL, Online RL), and GS-Pruning to reduce model size without proportionate capability loss.
- Mano-P supports local-only operation (no telemetry/cloud fallback) and optional offload to a physically local USB 4.0 inference accelerator to preserve data sovereignty.
Connected Companies & Entities
5 Entities mappedOntology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
Mano-P: Open-Source On‑Device GUI Agent; Apple CEO Change
The article introduces Mano-P, an open-source, on-device GUI Agent for macOS that uses a pure-vision approach to operate graphical interfaces. Mano-P is released under Apache 2.0 (GitHub: Mininglamp-AI/Mano-P) and provides a three-stage training pipeline (Supervised Fine-Tuning, Offline RL, Online RL) plus a think-act-verify inference loop for self-correction. Benchmarks report a 58.2% success rate for Mano-P's 72B model on OSWorld and a 41.7 NavEval score on WebRetriever Protocol I, outperforming several contemporaries. A quantized 4B w4a16 model is shown running locally on Apple M4 Pro (476 tokens/s prefill, 76 tokens/s decode, 4.3 GB peak memory). The piece also notes Apple announced Tim Cook will step down to Executive Chairman with John Ternus becoming CEO on September 1.
Mano-P On-Device GUI Agent, Cider SDK, Mano-AFK Open-Source
Mininglamp-AI has open-sourced a full on-device GUI agent stack for Apple Silicon that includes the Mano-P 1.0-4B local model, the Cider INT8 activation quantization inference SDK, and Mano-AFK (an end-to-end automated app builder). The stack is designed to run entirely offline so screenshots and task data remain on-device. Mano-P uses a three-stage training pipeline (SFT → Offline RL → Online RL) and a think-act-verify loop; benchmarked results are reported for larger Mano-P variants. Cider provides W8A8 and W4A8 modes via custom Metal kernels to enable INT8 activation quantized inference on MLX, claiming measurable prefill speedups versus MLX native modes. The release includes hardware guidance, performance figures, GitHub repos, and downloads on HuggingFace and ModelScope. Publication date: 2026-05-06.
Vision-Only AI Agents Break the Browser Boundary
The article describes a vision-only approach to GUI automation that lets AI agents interact with any on-screen application by reasoning over pixels rather than relying on browser DOM or OS accessibility APIs. It compares three approaches (CDP/HTML parsing, accessibility APIs, and vision-only), explains strengths and limitations, and presents Mano-P — an open-source, Apache‑2.0 GUI-aware agent model from Mininglamp-AI — as a working vision-first system. Mano-P reportedly achieves a 58.2% success rate on the cross-application OSWorld benchmark (vs. 45.0% for the runner-up), outperforms competitors on a web navigation benchmark (41.7 NavEval), and runs on-device using a 4B-parameter model with w4a16 quantization. The piece covers training (SFT, offline RL, online RL), model compression (GS-Pruning), edge performance metrics on Apple M4 Pro, and a three-phase release plan (skills released; local models/SDK and training methods forthcoming).
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.