2026 年如何用 OGAM 為 Android 手機挑選本機 AI 模型：Fast、Balanced 或 Best

You want a local AI model that runs well on your Android phone, and you do not want to read quantization tables to find it. Off Grid AI Mobile (OGAM) now opens the Models screen with three picks worked out for your phone's RAM: Fast, Balanced and Best. Tap one, it downloads, and it becomes the model your chats use. After the download, the model runs on your phone's CPU, GPU or NPU, and nothing you type is sent anywhere.

The same three picks are on the Image tab for image generation and on the Speech tab for voice transcription.

What you need

An Android phone. The picks scale with RAM, so a 6 GB phone and a 16 GB phone see different models.
OGAM 0.0.116 or later. The picks arrived in 0.0.116 and are in the current stable release, 0.0.117. Download OGAM for Android.
Wi-Fi for the download. Chat models are hundreds of megabytes to a few gigabytes. After that, no connection is needed.
No account and no payment. The model picks are part of the free app.

How OGAM chooses Fast, Balanced and Best

The picks come from one rule that lives in the app's shared model catalog. It reads your phone's total RAM and each model's download size, then keeps only the models that run well on your phone.

An Android phone does not give one app all of its RAM, and a model needs more memory than its file size while it runs (weights, plus the cache that holds your conversation). The app's phone budget is measured from that: a file under about 36% of your RAM runs well, a file between 36% and 50% is marked Tight, and anything larger is not offered. On a 12 GB phone, that is under 4.3 GB to run well and up to 6 GB at most.

From the models that run well:

Best is the highest-quality chat model in the catalog that fits.
Balanced is the next best.
Fast is a model that runs at the speed of a 4B model or faster, and faster than Balanced. A mixture-of-experts model counts by its active parameters, so a large model with a small active part can be the Fast pick.

Picks come from model makers people know, such as Qwen, Gemma, Llama, Mistral and Phi. A model that a newer one replaces is never picked, and neither is a model your app's engine is too old to load. As an example from the release notes, a 12 GB phone gets Qwen 3.5 0.8B, 2B and 4B.

On Android, when you run on the CPU or the NPU, the app downloads the Q4_0 version of a model if one exists, because it runs faster there.

Step 1: Open the picks

Open OGAM and tap Models in the bottom bar.
Stay on the Text tab. At the top, under For this phone, you see three small cards side by side: Fast, Balanced and Best.
Each card shows the model name and its download size.

Step 2: Download one and switch to it

Tap the download icon on the card you want.
The card shows the download progress as a percentage.
When the download finishes, the model becomes your active chat model. The card gets a highlighted border and a chat icon.
Tap the chat icon (Chat now) to open a chat with that model.

If you already downloaded a pick earlier, its card shows a circle icon instead. Tap it to switch to that model without downloading again.

If you download a model from the list below the picks instead, the app also makes it your active model when it finishes and shows "[model name] is ready" with Later and Chat now.

If you switch to a model that does not fit in the memory that is free right now, the app keeps your current model loaded, so you do not lose your chat.

Step 3: Pick an image model and a speech model

Tap the Image tab. The same three cards appear under For this phone. On this tab they take the place of the old RAM banner.
Tap the download icon on a card. When it finishes, the app says the model is ready, and the pick becomes the image model your chats use.
Tap the Speech tab. The picks here are multilingual speech-to-text models only, so they work whatever language you speak.

Image and speech models have no quality score in the catalog, so these picks go by size: Best is the largest that runs well, Fast the smallest with at least 2 billion parameters, and Balanced the one in between.

Step 4: Choose a size in plain words

When you tap a model in the list to open its details, the Sizes section shows up to three choices instead of quantization names:

Recommended: the file that suits your phone best. It is listed first.
Smaller: the smallest file below the recommended one, for when storage or memory is short.
Best: the largest file above the recommended one, for the highest quality your phone can run.

Every file is still there. Tap All files under the list to see them all, and Fewer sizes to go back. Files too big for your phone are hidden, and files that only just fit carry a Tight label.

Step 5: Get new models without an app update

New models are added to the catalog between app releases. To get them:

On the Models screen, tap the refresh icon at the top right, next to the downloads icon.
A sheet opens. It names each new model, for example "2 new models", and the new models appear in the list straight away.
If nothing changed, the sheet says Up to date. If your phone is offline, it says it could not check and asks you to check your connection.

The picks are worked out again with the new models included, so a better model for your phone can move into the Best slot.

What this does not do

It does not test speed on your phone. The picks are based on RAM, file size and the catalog's quality ranking, not a benchmark run on your device. To tune speed, use Use fastest settings on the Model Settings screen, which tests each backend your phone can run in 30 seconds or less.
It does not look at free storage when it picks. The app checks free space when the download starts and tells you how much it needs if space is short.
It does not pick from all of Hugging Face. The picks come from the curated catalog. You can still search Hugging Face from the Text tab, and the search hides files your phone cannot run.
A slot can be empty. On a phone with little RAM, there may be fewer than three models that run well, and the row shows only the slots it can fill.

FAQ

What is the best local AI model for an Android phone with 8 GB of RAM?
Open the Text tab in OGAM and look at the Best card: that is the highest-quality catalog model that runs well in 8 GB. On 8 GB, files under about 2.9 GB run well, so the picks are small models at 4-bit.

Which is faster on Android, Fast or Balanced?
Fast is always the quicker of the two. It runs at the speed of a 4B model or faster, while Balanced is the next-best quality model and can be larger.

Does picking a model need internet?
Only for the download and for checking for new models. After the download, chat, image generation and transcription with that model run on the phone with no connection.

Why does a model show Tight?
Its file is between about 36% and 50% of your phone's RAM. It can load, but there is little room left for other apps and a long conversation, so a smaller size may be steadier.

Can I still choose a different model?
Yes. The list below the picks has the full curated catalog, and you can search Hugging Face. The picks are a starting point, not a limit.

Why did my picks change?
The picks follow the catalog. After you tap the refresh icon and new models arrive, a newer or better model that fits your phone can take a slot.