DeepSeek for On-device and Consumer Electronics — the five layers you actually need
Running DeepSeek on a phone. This is the highest-download direction of all — a single quantized GGUF has passed 2.7 million downloads.
Models
50
Toolchain
9
Reference impls
13
Hardware
85
1. Models
2. Toolchain
| Project | ★ | What it does |
|---|---|---|
| dsh-full-remote | 51 | Auditable, token-gated DeepSeek Harness remote gateway: mobile QR access, per-device sessi |
| coral-pocket-ai | 3 | 🌊 珊瑚口袋 AI (Coral Pocket) - 极致轻量的单文件 (Single-File) 跨平台 AI 聚合客户端。原生移动端体验 (Mobile-First PWA)， |
| Mobile-Shrimp | 1 | Mobile chat bot adapters that bridge WeChat and Feishu messages to local Codex or DeepSeek |
| laravel-mtn-momo-ai | 1 | A powerful Laravel package that integrates MTN Mobile Money API with AI capabilities, prov |
| Multimodal-Industrial-Agent | 1 | 基于 YOLOv8/ONNX + SciPy/FFT + DeepSeek API 的多模态端侧 AI Agent 巡检系统 |
| harness-ai-app | 0 | Mobile client for Harness AI — follow, answer and approve your hosted agent sessions from |
| Smart-Finance-Platform | 0 | Full-stack M-Pesa financial intelligence platform & mobile API gateway. Features automated |
| dsh-plugin-deepseek-web | 0 | 将 DeepSeek 网页版（`https://chat.deepseek.com`）接入 DSH 桌面端侧边栏的插件。 在左侧工作区上方加了一个 "DS" 图标，点击后用 El |
| litikj-agent-framework | 0 | 面向未来的端侧 AI 智能体微内核框架 (Local-First AI Agent Operating Framework) | 极简底座，支持 0-Token 快速直达、零摩擦环 |
3. Reference implementations
| Project | ★ | What it does |
|---|---|---|
| DSH-better-sidebar | 4,913 | 开放的侧边栏底座，支持三方拓展注册新侧边栏页面。内置文件渲染编辑/终端/侧边对话/Git/子代理页面 ｜ Open sidebar foundation, supports thi |
| local-code-community | 115 | Source code of Local Code, a private AI agent for Mac, iPad and iPhone. Runs DeepSeek 4.1 |
| CyberBrain | 16 | CyberBrain is an advanced AI project designed specifically for training artificial intelli |
| deepseek-harness-for-vscode-unofficial | 2 | 非官方 DeepSeek Harness VS Code 客户端：侧边栏嵌入官方 Web UI。Unofficial third-party VS Code client for |
| dsh-balance | 2 | 在 DeepSeek Harness 桌面端侧边栏常驻显示 DeepSeek 账号余额的插件（宿主统一取数 + 浏览器浮层） |
| dsh-slamdunk-skin | 1 | Plugin and Skin Collection for DeepSeek Harness (DSH) Web UI An extension suite for DSH We |
| Jarvis | 0 | Jarvis is a mobile voice assistant built with Flutter. It integrates speech-to-text, text- |
| YueDong | 0 | A sports e-commerce platform for college students, featuring a storefront, user center, ad |
| sol-pi-engram-evolution | 0 | SoL-Pi × DeepSeek Engram 融合系统 — 端侧 AI 智能体 |
| ScheutieOS | 0 | DeepSeek 网页的独立窗口客户端，侧重长对话导出（PDF / Markdown）与字数统计 |
4. Hardware
What runs on what. These are working examples, not vendor claims.
| Project | ★ | Hardware target |
|---|---|---|
| dsh-Go | 10,925 | dsh-go — DeepSeek Harness 装进口袋：手机本机运行，也能遥控你的电脑。 |
| rikkahub | 8,661 | RikkaHub is an Android APP that supports for multiple LLM providers. |
| DSHA | 3,043 | 免 ROOT 免 Termux，在手机上跑 DeepSeek Harness。完整 Ubuntu 环境 + proroot 零 ptrace 开销 · AI 输出实时上屏 · AD |
| maid | 2,755 | Maid is a free and open source application for interfacing with llama.cpp models locally, |
| gpt_mobile | 1,268 | Chat app for Android that supports answers from multiple LLMs at once. Bring your own API |
| DSH-Android | 828 | 免 ROOT 免 Termux，一个 APK 在安卓手机上跑 DeepSeek Harness。引擎内嵌 App 进程 · 读屏操作 App · 改系统设置 · 定时/条件自动化 |
| oriveo | 554 | Open-source, bring-your-own-key (BYOK) multi-model AI chat client for iOS, Android and the |
| ai-companion-kit | 552 | 给你的 AI 一副身体，和一点人味 —— 会记事、会看场合的 QQ 陪聊 + 能控制安卓手机的 MCP |
| ZeroClaw-Android | 308 | Run AI agents 24/7 on your Android phone. Native Rust core, 25+ providers (OpenAI, Claude, |
| on-device-vision-model | 203 | On-device Vision Model - a fully offline, free vision pipeline that gives text-only models |
5. Pitfalls
- Phone RAM is the hard limit. An 8GB phone has 3-4GB usable; a 7B model at Q4 is about 4.5GB and will not fit.
- Heat and battery drain are hidden costs. Sustained inference throttles the SoC — real experience beats benchmarks.
- Running a full Linux environment on Android without root relies on proot; compatibility varies by handset.
- Quantization loses accuracy: Q4 is fine for classification, clearly degraded for reasoning.
- App stores limit download size — do not ship a 4GB first launch.