Although there are many AI chatbot apps in the App Store, they all need to be connected to the Internet to use them. For some users, they may be worried about privacy issues. This article will recommend an open source free App “fullmoon: local intelligence” that can run LLMs language models locally. It currently supports the latest 3b and 1b models of Llama 3.2. It can be downloaded and used directly in the App without installation or any troublesome settings. The operation is very simple.
fullmoon: local intelligence, a free app for running LLMs locally and offline on iPhone
fullmoon is a free and open source App. In addition to iPhone, it also supports macOS and visionOS. People who need it can also try it.
Currently, the only models provided are the 3b and 1b models of Llama 3.2. For part 3b, I tested the conversation in Chinese with no problem, but the generation speed will be a little slower, but the length of the answer is much longer than expected.
After opening fullmoon for the first time, you can select the model you want to download (now the DeepSeek 1.5b model is added). If your iPhone is not too old, such as iPhone 13, I tested llama-3.2-3b-instruct-4bit and it was no problem. Although the response speed is a little slower, it is acceptable. 1b will definitely be faster, but with fewer parameters, the quality of the responses will be slightly worse:
After the model is downloaded, you can start chatting with the AI:
I also deliberately turned on airplane mode and turned off WiFi to test whether it was really running locally. The result was that, even if there is no Internet, I can chat with the AI:
Although compared with some larger language models, the parameters of 3b are still too few, but I think the quality of the responses is already very good, not to mention that I require answers in Chinese. However, it can still be noticed that some words used by AI will suddenly change to English, such as asking for help at the seventh point on the right side of the picture below. In the end, it should be asking for their help. As a result, AI becomes ABC and the answer is asking for their help. The eighth point is also, suddenly practice sell-care in English. But this does not affect the use, you can still understand it:
What surprised me was that when I simply asked the question “How to solve money anxiety”, llama-3.2-3b-instruct-4bit could also answer a lot of content. Generally, most of these small models only reply with a few sentences.
Click the function icon in the upper left corner to view all chat records and open a new chat room window from here. The model section can also be switched and downloaded other new ones:
In addition, in the settings, you can also adjust the appearance. If you don’t like the default one:
Editor’s note: Recently, Fullmoon has also added the DeepSeek R1 1.5b small model, which is smarter than Llama 3.2 in some aspects of performance. It also has a thought process:
It is a pity that this app does not provide the function of adding prompt words (Promp), and it may be added later.
This App is open source software, available in GitHub Share the source code so that people in need can take a look.
Source: KOCPC Chinese