Under the Hood

Earthscale runs on llama.cpp, compiled directly into the app as a native framework. Your models are standard GGUF files, an open format used across the community.

The engine runs on the Metal GPU. Every layer of the network, through Metal, not the CPU. That’s the difference between an iPhone that gets warm and one that gets hot.

Every model in Earthscale is an open-weight model, published by its original creator, such as Qwen and ggml-org. We never fine-tune or modify them, and each one comes with its own license. We don’t train models. We choose great ones, and make them run beautifully on iPhone.

Models come in different levels of precision, from Q2_K to Q8_0. On an iPhone with less memory, a tier can automatically use a lighter version of the very same model. Same weights. Fewer bits. You never choose a level. Your iPhone does.

Earthscale understands images with mtmd, llama.cpp’s own multimodal library. The vision encoder and the language model are trained together, as one, and we use them exactly as their authors shipped them.

On our smallest tiers, two models work in sequence. One looks at your photo and describes it. The other reasons about that description. Two steps. One seamless answer. Your photo is processed once and remembered, so reasoning never starts over.

Reasoning Mode works in six stages: analyze, plan, solve, critique, revise, finalize. Each stage is its own pass over the same model, with its own budget to think.

And it knows how hard to think. Earthscale reads your question first, its length, its steps, its math, and decides how deep to go. A simple question gets a quick answer. A hard one gets the full treatment. The heaviest tier can use up to nine passes, with more room to work.

Earthscale remembers across conversations, and it remembers well. For every memory, it weighs four things: what it means, which words matter most, how recent it is, and which project it belongs to. Then it ranks them, removes the duplicates, and brings the best into your conversation.

All of it happens on your iPhone. Nothing about what Earthscale remembers, or how it decided, ever leaves.

Earthscale keeps the model’s working memory of your conversation, the KV cache, and adds only what’s new.

When a conversation grows long, the oldest messages become a running summary. Nothing is cut off. Nothing overflows.

Privacy is built into the architecture. No URLSession. No sockets. From the moment you hit send to the moment a word comes back, there is no path out.

Earthscale uses the network for two things only: downloading a model you chose, and your App Store subscription. Both happen outside the engine.