01 · MODELS
Which AI model is behind MATE?
A fair question, and one a growing number of Swiss providers leave unanswered. Here is ours: which models we run, where they run, and what your device needs to run them itself.
MATE Cloud
When a request needs more than your device can carry, it goes to our own infrastructure in Switzerland.
Our own data centres
The models run on hardware Swisswise owns and operates in Switzerland. There is no hyperscaler underneath, which is what makes the processing location a commitment rather than a contract term someone else could revise at renewal.
Open weights, several families
We use curated and in part modified language models with open weights, among them the Qwen and Gemma families, plus a separate model for image generation. Open weights mean the model is published and inspectable — we run it ourselves instead of buying it through someone else's interface.
No third-party AI interfaces
Your requests are not passed on to an external AI service. AI processing stays in our data centres, and nothing you write is used to train a model. The request is processed and then not kept: once the answer is written, none of its content remains on our machines. Your history lives on your device, not with us.
02 · ON YOUR PHONE
Models that run on the device itself. Offline.
On-device Mode answers without any data centre at all. Nothing ships inside the app: you choose a curated and tested model, MATE downloads it, and it stays on your device.
iPhone and iPad
| In the app | Model | Recommended devices |
|---|---|---|
| Sprite | Qwen 3.5 0.8B · 0.9 GB download | iPhone 15 Pro or an iPhone 16 and newer. On iPad: Air from 2022, mini from 2024, any iPad Pro with an M-chip |
| Companion | Qwen 3.5 2B · 1.5 GB download | Same devices as Sprite. If you attach photos often, a 12 GB iPhone or an M-series iPad has more headroom |
| ScholarExperimental | Gemma 4 E2B · 3.6 GB download | iPad Pro with 1 TB or more, iPad Air M4, iPad Pro M5. It can also run on iPhone Air and iPhone 17 Pro |
| AtlasExperimental | Gemma 4 E4B · 5.2 GB download | An iPad model: iPad Pro with 1 TB of storage or more, because those versions carry 16 GB of memory |
| VoiceBeta | Marvis TTS 250M, for spoken answers | The same devices that can run a chat model. On its own it is of little use, so we recommend it only alongside one |
Android
| In the app | Model | Recommended devices |
|---|---|---|
| Companion | Qwen 2.5 1.5B Instruct · 1.6 GB download | A phone with at least 8 GB of memory: Pixel 8 and newer, Galaxy S24 and newer, Fairphone 6, Nothing Phone (2) and newer |
| Scholar | Gemma 4 E2B · 2.6 GB download | A phone in the 12 GB class: Pixel 9, 10 or 11, Galaxy S24+ and S24 Ultra, Galaxy S25 and newer, Nothing Phone (3) |
| Atlas | Gemma 4 E4B · 3.7 GB download | Flagships with 16 GB: Pixel 9 Pro and 10 Pro, Pixel 11 Pro from 512 GB up, Galaxy S26 Ultra with 1 TB, Nothing Phone (3) with 512 GB. Where two storage tiers exist, the larger one is the right choice |
These are recommendations, not a compatibility list. MATE measures the memory your device actually makes available before it offers a download, and says so when a model will not fit — that check, not this table, is the authority.
03 · ON YOUR MAC
Larger models on Apple Silicon
For every Apple Silicon Mac there is a fitting model, curated and tested. The range starts small enough for an entry-level machine and goes up to models that would be out of reach on a phone. The more memory your Mac has, the larger the model you can use.
| In the app | Model | Recommended devices |
|---|---|---|
| Tiny Test Pilot | Qwen 3.5 0.8B | Any Apple Silicon Mac. On an 8 GB Mac, close other apps first |
| Pocket Sprint | Qwen 3.5 2B | Any Apple Silicon Mac. On an 8 GB Mac, close other apps first |
| Everyday Local BrainDefault | Qwen 3.5 4B | 16 GB — the standard configuration of every Mac Apple currently sells |
| Gemma Quick Draft | Gemma 4 E2B | 16 GB |
| Smooth Writer | Gemma 4 E4B | 16 GB |
| Sharper Thinker | Qwen 3.5 9B | 24 GB for a comfortable experience |
| Premium Writer | Gemma 4 12B | 32 GB. Available on MacBook Air and Pro, and on the four-port iMac |
| Local Flagship | Qwen 3.6 35B A3B | 64 GB. MacBook Pro with an M4 Pro, M4 Max, M5 Pro or M5 Max, or a Mac Studio |
| Gemma Flagship | Gemma 4 26B A4B | 64 GB, same machines as above |
| Bonsai Image Studio | Bonsai Image Ternary 4B, for local image generation | 24 GB. Image models briefly need considerably more memory while computing than they do at rest, so this recommendation sits above the figure shown in the app |
Memory decides, not the model year — a 2021 Mac with 32 GB carries more than a current one with 16. What Apple offers as a configuration changed several times during 2026, so check the memory rather than the name.
What you can check for yourself
Transparency that cannot be verified is just a nicer way of asking for trust.
- Every local model is downloaded on demand and verified against a checksum recorded in our source code before it is loaded. A file that does not match is not used.
- On the Mac, the local runtime works with network access switched off once it has started. It answers from the model on your disk, not from somewhere else.
- Local answers reach no data centre — not ours either. You can see it: turn off the network and keep chatting.
- The open weights we build on are published by their originators and can be inspected by anyone, independently of us.