100% 在本机跑 · 无遥测 · 无账号 100% on-device · no telemetry · no account

在你的手机上,
完全离线地跑 AI
Run AI entirely offline
on your own phone

一个 Android 本地 AI 应用:与大语言模型对话(可发图片、文件与文档)、画图、给图片反推标签,还能通过插件扩展能力。模型在你设备上推理,数据不出手机。 An Android app that runs AI locally: chat with LLMs (images, files and documents welcome), generate images, caption photos, and extend it with plugins. Inference happens on your device — your data never leaves it.

版本Version 2.0.0 体积Size 61 MB 系统OS Android 9+ 架构ABI arm64-v8a MIT 许可code
能力Capabilities

整套本地 AI 技术栈A whole local AI stack

每一块都在手机上跑。没有云端账户,也不需要联网 —— 除了你自己去下载模型的那一刻。 Every piece runs on the phone. No cloud account, and no network needed — except the moment you fetch a model yourself.

多模态对话Multimodal chat

双后端:LiteRT-LM 与 llama.cpp。可发图片、文件和文档,思考过程与正文分开显示,支持中断与继续。 Two backends: LiteRT-LM and llama.cpp. Send images, files and documents; reasoning is shown apart from the answer, and you can interrupt and resume.

.litertlm.ggufmmproj

本地绘图On-device drawing

内置 stable-diffusion.cpp。文生图与图生图、可选 Vulkan GPU 加速、TAESD 加速解码,还能看逐步生成的过程预览。 Bundled stable-diffusion.cpp. Text-to-image and image-to-image, optional Vulkan GPU, TAESD for faster decoding, plus a step-by-step progress preview.

SD 1.5 / SDXLVulkanLCM-LoRA

图像打标Image tagging

离线 ONNX 反推标签(WD14 系)。标签能一键送到文生图或图生图;打标模型也能被看不见图的对话模型当工具调用。 Offline ONNX tagging (WD14 family). Send tags straight to text-to-image or image-to-image; a chat model that can't see images can call the tagger as a tool.

ONNXWD14NNAPI

把手机当服务器Use the phone as a server

把已加载的模型开成 OpenAI 兼容接口,电脑上的客户端或 Agent 都能连。对话与绘图各一套,端口独立、可只开其中一个。 Expose loaded models over an OpenAI-compatible API so desktop clients and agents can connect. Chat and drawing are separate services with their own ports.

:8080 chat:8081 draw

插件式扩展Plugin extensions

语音合成这类能力放在独立 APK 里,不必都塞进主程序。协议公开,第三方插件同样能用,你也可以自己写。 Capabilities like speech synthesis ship as separate APKs instead of being crammed into the app. The protocol is public — third-party plugins work, and you can write your own.

Ponko-EXplugin API v12

隐私是设计而不是选项Privacy by design, not a setting

没有账号、没有自动遥测、没有默认联网。模型、会话、图片都留在应用私有目录里 —— 想分享什么,你自己决定。 No account, no automatic telemetry, no default network access. Models, chats and images stay in the app's private storage — what leaves is up to you.

离线offline无追踪no tracking
扩展Extend

能力可以长在插件里Capabilities can grow as plugins

插件是独立 APK,跑在自己的进程里:它崩了不会拖死 Ponko。宿主负责渲染界面,插件只提供数据 —— 所以深浅色、圆角、布局都是统一的。 A plugin is a separate APK in its own process: if it crashes, Ponko survives. The host renders the UI and the plugin only supplies data, so theming and layout stay consistent.

Ponko-EX · 插件开发仓库Ponko-EX · plugin development repo

中英双语的插件开发手册、示例插件,以及第一个发布版 VoxCPM2 本地语音合成。也可以直接拿它当自己插件的起点。 A bilingual plugin guide, sample plugins, and the first release: VoxCPM2 local speech synthesis. Use it as the starting point for your own.

打开 Ponko-EXOpen Ponko-EX
许可License

代码开源,美术资源保留Open code, reserved artwork

这个仓库不是单一许可证 —— 代码与美术资源分开授权。 This project is not single-licensed — code and artwork are covered by different terms.

源代码Source code
MIT —— 可自用、修改、商用、闭源分发,保留版权声明即可,不提供担保。 MIT — use, modify, ship commercially, keep it closed; just keep the notice, and no warranty.
美术资源Artwork
应用图标、角色立绘、原始画稿 —— 版权归作者所有,不适用 MIT。 App icon, character portrait, original drawing — copyright the author, not MIT.
模型权重Model weights
不随应用分发,各自遵循来源许可;需要你自备。 Not bundled; each follows its source licence, and you supply your own.
第三方组件Third-party parts
stable-diffusion.cpp、llama.cpp、LiteRT-LM、ONNX Runtime 等各循其许可证,详见 NOTICE。 stable-diffusion.cpp, llama.cpp, LiteRT-LM, ONNX Runtime and others under their own licences — see NOTICE.