A chat app that keeps the conversation on your own machine.
Gadeon is a chat application for the Mac and the iPhone. The difference is where the thinking happens. There is no server behind it, so when you type a question, nothing is uploaded and nothing is logged. The answer is computed by your own hardware, a few centimetres from your hands.
It works on a plane, on a train, in a basement, and on the day the service you were relying on shuts down. Once the model is downloaded, the network is optional.
The same app on both, in whichever theme the system is wearing. Every word on these two screens was produced by the machine it is displayed on.
It reads, looks, listens, and shows its work.
-
01
Talk or type
Ask by keyboard, or tap the microphone and speak. Replies can be read back aloud in a voice that is also synthesized on the device.
-
02
Show it a picture
Attach a photo and ask about it. Some models take sound and video too, and you can watch the frames go by as it looks.
-
03
Hand it a document
Drop in a PDF, a Word file, a spreadsheet or a page of notes. It reads the text out of them, tables included, and answers from what is inside.
-
04
Watch it think
The stronger models reason before answering, and you can read that reasoning, cut it short when you have seen enough, or switch it off for a faster reply.
-
05
Give it tools
A calculator it can actually use, the clock, the weather, the headlines, Wikipedia and web search. Each one is a switch you control.
-
06
Keep the result
Answers are typeset properly, mathematics and all. Any conversation can be searched, renamed, shared as a PDF, and recovered from the trash for thirty days.
Nowhere, unless you flip a switch.
Most assistants send your sentence to a company. Gadeon has nowhere to send it. The whole model sits in your device's memory and runs on its own chip, so privacy here is not a policy that could be revised. It is the shape of the thing.
No account. No sign-in. No subscription. No analytics and no telemetry.
Small and quick, or large and thoughtful.
Several open models are offered, and you choose. A small one answers almost instantly and fits in a phone. A large one is noticeably smarter and wants a Mac with memory to spare. Each is downloaded once and then belongs to you.
The app only offers what your device can actually hold, so the list you see is shorter than this one and always runnable.
Measured against llama.cpp on the same machine.
Tokens per second on a MacBook Air (M3, 24 GB): reading a 512-token prompt, then writing 128 tokens. The second number in each cell is llama.cpp on the same file and the same protocol. The last column is the same model writing with its own draft head, which most of ours carry.
| Model | Read | Write | Write, draftedDrafted |
|---|---|---|---|
| Qwen3.5 4B | 341 / 316 | 27.0 / 23.9 | 34.6 / 24.3 |
| Qwen3.5 9B | 186 / 179 | 15.6 / 14.1 | 18.2 / 16.8 |
| Qwen3.8 27B | 51.0 / 52.4 | 6.0 / 5.7 | 8.5 / 5.5 |
| Gemma 4 E2B | 577 / 679 | 49.8 / 53.6 | 65.1 |
| Gemma 4 E4B | 298 / 336 | 26.9 / 25.8 | 44.6 |
| Gemma 4 12B | 116 / 120 | 11.8 / 11.5 | 20.7 |
llama.cpp build b10796 on the pristine files; ours on the same weights with our own vision, audio and draft tensors added. llama.cpp runs no draft head for Gemma, so that cell is ours alone. One cooled run each, 2026-09-07.
And in a pocket, the same protocol in the app itself:
| Device | Model | Read | Write | Write, draftedDrafted |
|---|---|---|---|---|
| iPad mini A17 Pro | Gemma 4 E4B | 123 | 13.8 | 20.6 |
| iPhone 15 Pro | Gemma 4 E4B | 134 | 13.7 | 20.9 |
| iPhone 12 mini | Gemma 4 E2B | 93 | 16.2 |
The 4 GB phones write without the draft head on purpose: its extra weights do not stay resident there, and paging them every step is slower than not drafting.