How does on-device AI journaling work on iPhone?

Verified against Everglow 1.5 on 10 October 2026.

At a glance

Setting it up

During setup, Everglow offers three ways to run its AI: on this phone, Everglow Cloud, or your own API key. Choosing the phone starts the model download from Hugging Face. Here is what the download does.

• It uses Wi-Fi only, never cellular data.
• It keeps the screen from locking while it runs, and you can keep using the app. If you leave Everglow, it may remind you that the download finishes faster with the app open.
• If the connection drops partway, the app keeps what it already has and offers to resume. If the partial file can’t be used, it offers to start over.
• Before the model is loaded, the app checks the file against the SHA-256 hash of the exact revision it expects. A file that doesn’t match is deleted, not loaded.
• If Hugging Face doesn’t answer, the app tries a mirror, hf-mirror.com, held to the same check. If no download server answers, which the app notes is common in mainland China, it says so and offers DeepSeek, a cloud model, instead.

The model is kept in the app’s own storage, apart from your journal, so Everglow’s iCloud backup doesn’t include it. On a new iPhone you download it again. You can delete it in Settings to free the space.

How your iPhone’s memory matters

Setup checks how much memory the phone has. If there isn’t enough for the model, setup doesn’t offer on-device AI, says why, and offers Everglow Cloud or your own API key instead.

Memory also sets the context budget: how many tokens the model can hold at once, counting both what it reads and what it writes. Phones that report at least 6.5 GB of memory get 32,768 tokens; others get 24,576. Because the budget is shared, a very long entry or a long conversation leaves less room for the reply.

What runs on the model

• Reading each new entry. In the background, the model writes a short summary and picks out the feeling, the people, places and things mentioned, and the topics.
• Ask. Search finds candidate passages and the model keeps the relevant ones. It then writes the answer from a capped amount of evidence, about 6,000 tokens and usually no more than 12 entries, and shows the dated excerpts it used.
• The companion. Replies in chat draw on relevant entries. On the phone, the conversation stays loaded between turns instead of being re-read each time. In a very long chat, the oldest turns drop out of the model’s working memory and the recent ones stay.
• People pages. Facts about the people, places and things you write about, each linked to the entry it came from.
• Find the date. Working out when something you describe happened.

What doesn’t use the model

• Writing, editing, bookmarks and keyword search.
• The vectors behind search by meaning. Everglow computes them on the phone with Apple’s NaturalLanguage framework, not with Gemma.
• Drawing the relationship map. The people and ties on it come from earlier reading; arranging them is plain code.
• Markdown export, iCloud backup and app lock.

One request at a time

The phone runs one model request at a time. When you chat or ask a question, that request goes first; background reading waits, or stops and runs again afterward. A large import is read entry by entry. The app’s own note says a few hundred entries can take a while, and an import picks up where it left off if you quit the app. For a single import, you can choose another engine you’ve already set up.

Apple Intelligence isn’t involved

Everglow ships its own model and runtime. In version 1.5, none of its AI features use Apple Intelligence or Apple’s Foundation Models framework, so on-device AI in Everglow doesn’t depend on Apple Intelligence being available or turned on. It depends on the model download, the phone’s memory and free storage.

The trade-offs

• Speed and battery. On-device AI is slower than hosted AI and uses more battery. The app’s setup screen lists both as the cost of this option.
• Storage. About 2.6 GB for the model, on top of the app.
• Depth. It is a small model with a bounded context. Hosted AI is usually faster and handles longer, more detailed questions better.
• Accuracy. Any AI can get details wrong. Ask shows the dated excerpts it used, so you can check them against your entries.
• GPU or CPU. The model runs on the GPU through Metal unless you turn that off in Settings, where the app notes the GPU is faster on most devices. The change applies after you restart the app.

Privacy with on-device AI

With on-device AI, your entries are processed on the iPhone and nothing is sent anywhere for AI. The journal archive is stored on the phone whichever engine you use. Downloading the model sends no journal content. iCloud backup is a separate, optional choice. The first time you switch to Everglow Cloud or your own API key, the app asks for your consent and names the provider.

Read more

• Can an AI journal work offline?
• Where your writing goes with each AI engine
• Using your own API key
• Free limits, Unlimited and Cloud