How to use deepseek Offline on your iPhone
Fullmoon Installation
Search for fullmoon in the App Store, which is currently free. You can run the mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-4bit model directly on your iPhone.
Address: https://apps.apple.com/app/fullmoon-local-intelligence/id6727014156
Configuration steps are as follows:
Open Settingsinstall, and select
Note: I followed the rules you provided, preserving the original Markdown structure and content. Of course, there are many other apps that support this, some of which require payment, but I won’t mention them here.
Principle
After analyzing the principle, it appears to be implemented using Apple’s MLX Swift. You can directly modify the LLMEval example and make a few changes to run DeepSeek. If you’re an iOS developer, you should be able to implement it quickly.
The following image is an example of my test:
Step 1: Get the code:
Step 2: Add the following code:
Step 3: Modify the default model to DeepSeek and run it on a real device. I have tested it on an iPhone 16 Pro and it runs smoothly with 1.5B and 7B models, but I haven’t been able to download the 14B model (my phone’s storage is full).
Conclusion
With the breakthrough in computing power and algorithm optimization, deploying large models on terminals has overcome the technical threshold of landing. The real challenge is how to unleash their commercial potential through scenario-based innovation. In the future, terminal-side intelligence will present three major evolutionary directions:
Firstly, combining model lightweighting with heterogeneous computing will drive low-threshold AI empowerment in various industries;
Secondly, the natural advantage of privacy protection and real-time decision-making will give birth to revolutionary applications in sensitive fields such as medical diagnosis and financial risk control;
Thirdly, by combining multimodal interaction and contextual perception capabilities, terminal large models will reconstruct the mobile ecosystem and create a new paradigm for interactive services from passive response to active service.
When devices truly possess cognitive intelligence, business value will no longer be limited to cost reduction and efficiency improvement, but will enter a new era of intelligent enhancement through human-machine collaboration.
Finally, I have created an AI beginner’s learning group. If you’re interested in growing together, you can join the group by sending “join” or “AI” to this public account chat.
Note: I’ve kept the formatting and layout as close to the original as possible, but some minor adjustments were made for readability.
