Show HN: Off Grid: 本地设备AI网页浏览，工具 视觉、图像、语音——速度快3倍
Off Grid实现了本地设备上AI调用工具的功能，支持网页搜索、计算器、时间日期和设备信息等操作，并具备自动工具循环能力。该工具完全离线运行，无需云服务或API密钥，且在手机上可快速执行，量化模型在q4_0模式下处理速度提升3倍，支持多种KV缓存配置。
1 分•作者: ali_chherawalla•大约 2 小时前
九天前我在这里发布了Off Grid，你们纷纷出现——124分，66条评论，我当天就修复了报告的bug，还有让开源项目变得值得的反馈。<p>你们告诉我你们想要什么。现在我带来了成果：<p>你的AI现在可以使用工具——完全离线运行。<p>网页搜索、计算器、日期/时间、设备信息——并支持自动工具循环。<p>你的3B参数模型不再只是生成文本。它现在可以推理、调用工具并综合结果。<p>在你的手机上。无需API密钥。无需服务器。无需云功能。<p>那又怎样？这意味着“本地玩具”和“实用助手”之间的差距大大缩小了。<p>你不需要GPT-4来查找信息并给出答案。运行在你的Snapdragon上的量化Qwen 3 / SMOLLM3就能迅速完成。<p>速度提升3倍，支持可配置的KV缓存。<p>现在你可以选择f16、q8_0和q4_0三种KV缓存类型。在q4_0模式下，原本每秒处理10个token的模型现在可以达到30个token。应用甚至会提示你...
查看原文
Nine days ago I posted Off Grid here and you showed up - 124 points, 66 comments, bug reports I fixed same-day, and the kind of feedback that makes open source worth it.<p>You told me what you wanted. Here's what I shipped:
Your AI can now use tools — entirely offline.<p>Web search, calculator, date/time, device info — with automatic tool loops.<p>Your 3B parameter model doesn't just generate text anymore. It reasons, calls tools, and synthesizes results.<p>On your phone. No API key. No server. No cloud function.<p>So what? It means the gap between "local toy" and "useful assistant" just got dramatically smaller.<p>You don't need GPT-4 to look something up and give you an answer. A quantized Qwen 3 / SMOLLM3 running on your Snapdragon can do it in no time.<p>3x faster with configurable KV cache.
You can now choose between f16, q8_0, and q4_0 KV cache types. On q4_0, models that were doing 10 tok/s are hitting 30. The app even nudges you after your first generation: "Hey, you could be running faster." One tap.<p>So what? The #1 complaint about on-device AI is "it's too slow to be useful." That argument just lost a lot of weight. 30tokens/second on a phone is faster than most people read.<p>Live on both stores. No sideloading. No Xcode.<p>Off Grid is now on the App Store and Google Play. Install it like any other app. Your parents could use this.<p>So what? On-device AI just went from "cool weekend project for developers" to "thing normal people can actually try." That matters because privacy shouldn't require a CS degree.<p>What hasn't changed:
- MIT licensed. Fully open source. Every line
- Zero data leaves your device. No analytics. No telemetry. No "anonymous usage data."
- Text gen (15-30 tok/s), image gen (5-10s on NPU), vision AI, voice transcription, document analysis
— all offline
- Bring any GGUF model. Run Qwen 3, Llama 3.2, Gemma 3, Phi-4, whatever you want.<p>I'm building this because I believe the phone in your pocket should be the most private computer you own — not the most surveilled. Every week the models get smaller and faster. The hardware is already there. The software just needs to catch up.<p>If this resonates, a star on GitHub genuinely helps: <a href="https://github.com/alichherawalla/off-grid-mobile" rel="nofollow">https://github.com/alichherawalla/off-grid-mobile</a><p>I'm in the comments. Tell me what to build next.