更多第三方订阅支持
Response from developer
感谢您的五星好评和建议！我们已在下一版本中加入 OpenCode Go 订阅支持（设置 → 连接 → 云端），之后也会继续增加更多第三方服务。
I have been waiting for an app like this FOREVER
-Lets you use your own local AI LLM running on your phone (most do this)
-Has the ability to retain and recall memories (most other apps out the door already)
-Also lets the LLM tool call locally and search the internet for responses
Basically , a ChatGPT like experience but running locally on the phone. I’ve tested loads and no app whatsoever I could find does all these things together and well. I tried this app a few weeks ago and tool calls only seemed to work with the Apple Intelligence model at the time so I saw a great potential but I was missing using a more powerful model with tool calls. But now I checked it again, and what do you know tool calling works for other models and extra models were added! Finally, a local AI app that covers all the needs I had to have a truly usable local AI system! And it does SO much more. You can use a plethora of in built skills, load custom skill.md skills, use
/create apps (I haven’t even explored this to be honest), use it for agentic uses, use the attune function to taylor the AI even more. I’ve got a lot more to test clearly, but it works amazingly for what I do and the level of extra depth available there is just leagues above anything else I have found. It’s just too many things to mention in one review! Also, it just works apart from a couple minor bugs unlike so many other apps that just crash randomly or glitch. It’s great to see it being developed so actively too and I’m gonna be more than happy to pay for this when my trial runs out.
Response from developer
Thank you, this made our day. A local assistant that remembers, uses tools and can search the web is exactly what we set out to build, so it's great to hear it clicked for you. Since you haven't explored apps yet, try this: ask your model to "make me a [anything] in HTML5". A tip calculator, a habit tracker, a quiz on a topic you're learning. The model writes it and MyLLM turns it into an app you can open, save to MyLLMos and keep using, all on your phone. And if you can tell us about those couple of bugs, we'd love to fix them. The Feedback & Ideas app in the MyLLMos gallery comes straight to us.
Outstanding
Response from developer
Thank you, this made our day. You're right that it's a bit complex, and we'd like to fix that rather than rely on people already knowing their way around. If you have a minute, tell us the one thing that confused you most or took longest to find. You can send it from the Feedback & Ideas app in the MyLLMos gallery, or just update this review. We read every one, and the next few releases are about making things simpler, not adding more.
不卡死了，输入后无反馈
Response from developer
谢谢你写得这么具体。你遇到的卡死不是模型的问题，是我们的问题。 本机推理目前不是流式输出的：整段回复要全部生成完才会一次性显示，而生成期间会 占用几乎所有 CPU 核心，所以界面看起来像死机了，滚动和退出都没有反应。回复越长，卡住的时间越久 —— 换模型帮不上忙，你的测试结果是对的。 我们正在改三处：本机回复改成逐字流式显示；降低推理线程的优先级和数量，保证界 面始终能滚动、能退出；减少上下文缓存的内存占用，避免内存紧张时整个应用被拖住。 在更新之前，先用较小的模型（例如 Qwen 2.5 0.5B 或 LFM2 1.2B），并把设置里的「最大回复长度」调小，卡顿会明显减轻。给你造成不便，很抱歉。 Thanks for the detail — the freezing is ours, not the model's. On-device inference isn't streamed: the whole reply must finish before anything appears, and generation takes nearly every CPU core, so the UI looks dead and won't scroll or quit. Longer replies, longer freeze, switching models doesn't help, your testing was right. We're changing three things: stream on-device replies token by token; lower the priority and thread count of inference so the UI stays responsive; cut the context cache's memory footprint. Until then, a smaller model plus a lower "Max response length" will noticeably reduce it. Sorry for the trouble and thank you for the report, we are fixing
Save your money this app is not it
Response from developer
Thank you for taking the time to write this. You're right about more than we'd like. Looking into your review, we found that on-device chats re-process the whole conversation on every message, which is what makes the phone run hot and slows long chats down. Long conversations could also overfill the model's memory, and the app didn't back off when the device got warm. We're fixing all three in the next update, along with a smaller memory footprint and larger models for the newest iPhones. Apple Intelligence is already built in if you'd like to try it in the meantime. We'd be glad to hear from you again once the update lands.
The apps on this are insane!
Love the idea
Response from developer
Thank you for bringing this up to our attention - and you've found a real bug. Hermes 3 3B was locked behind a RAM check that asks for 9GB, which no 8GB device passes, including your M3 iPad Air. That was never the intention; the description even says iPhone 15 Pro or newer. The next update relaxes that check so Hermes is selectable on your iPad. On bigger models like Qwen3.5 4B you're right that the built-in list has fallen behind. The on-device engine is on an older build of llama.cpp that predates those architectures, so adding them means upgrading the engine first. That work is underway, and 4B-class models are the reason for it. In the meantime do not hesitate to reach out directly for discussing features and the app itself i.e. MyLLM@opticell-limited.com - Thanks for taking the time to write it up.
经常卡死，输入后无响应
Becky
I especially like the real-time updates and alerts, it makes the whole experience feel dynamic and actually helpful in day-to-day use. There’s been no crashes, no weird bugs, just a clean and reliable experience.
Definitely worth downloading if you want something smart, responsive, and well put together.
Response from developer
Thank you ! Genuinely good to read. The real-time updates were the hardest part to get right without tipping over into noise, so it's nice to hear they land the way they were meant to. There's an update on the way that brings noticeably more capable models running entirely on your device. Thanks for taking the time to write this up.
