展示HN:适用于iOS的设备端(离线)人工智能SDK(大语言模型、视觉和稳定扩散)
嗨,HN,我正在构建Kuzco,这是一个用于在Apple设备(iPhone、iPad、Mac、Vision Pro)上本地运行AI模型的Swift SDK。我们的目标是让“离线 + 私密AI”像普通依赖项一样,能够轻松集成到SwiftUI/ UIKit应用中,而无需搭建服务器或支付按令牌计费的API费用。
它的功能包括:
- 文本生成与令牌流式传输(聊天风格用户体验)
- 视觉AI(本地分析图像)
- 通过Stable Diffusion生成图像
- 模型管理器,用于下载和管理设备上的模型
开发者体验:
- 创建与模型的会话
- 在生成过程中流式传输令牌
- 可选获取完整响应(主页上有代码示例,展示了具体流程)
我发布这个内容的原因是希望能从iOS开发者(以及任何进行设备推理的开发者)那里获得反馈:
- 什么会让这个SDK成为你“必用”的选择
- 你希望首先支持哪些模型(小巧、快速、够用 vs 较大、较慢、更好)
- 你目前面临的最大痛点是什么:性能、模型下载、用户体验、内存、应用大小、安全性等。
如果你想在准备好广泛分享时获得更新或提前访问,可以在网站上加入等候名单(无需信用卡)。
<a href="https://kuzco.co/" rel="nofollow">https://kuzco.co/</a>
查看原文
Hi HN, I’m building Kuzco, a Swift SDK for running AI models locally on Apple devices (iPhone, iPad, Mac, Vision Pro). The goal is to make “offline + private AI” feel like a normal dependency you can drop into a SwiftUI/UIKit app, without standing up servers or paying per-token API fees.
What it does - Text generation with token streaming (chat style UX) - Vision AI (analyze images locally) - Image generation via Stable Diffusion - Model manager to download and manage on-device models<p>Developer experience - create a session with a model - stream tokens as they generate - optionally get a full response (There’s a code snippet on the homepage showing the exact flow.)<p>Why I’m posting I’d love feedback from iOS devs (and anyone shipping on-device inference) on: - What would make this a “must use” SDK for you - What models you’d want supported first (small, fast, good-enough vs bigger, slower, better) - What your biggest pain is today: performance, model downloads, UX, memory, app size, safety, etc.<p>If you want updates / early access when it’s ready to share broadly, there’s a waitlist on the site (no credit card).<p><a href="https://kuzco.co/" rel="nofollow">https://kuzco.co/</a>