Is Your Phone Fast Enough for Local AI?

Minimum Requirements

Before you download TokForge, check that your phone meets these baseline specs:

• Android 8.0+ (API 26 or higher) on a 64-bit ARM processor
• 4GB RAM: Minimum; small models only (around 1B parameters)
• 6GB RAM: Runs 2 to 3B models
• 8GB RAM: Recommended; runs 4B models comfortably
• Free storage for each model: from about 0.6GB for the smallest to about 21GB for the largest

If your phone has less than 4GB of RAM, local AI may not be practical. Consider using the Remote API backend to connect to a server instead.

RAM Tier Guide

The amount of RAM in your phone directly determines which AI models you can run and how fast they'll respond. Here's a quick reference:

Chipset Matters Too

RAM tells only part of the story. The processor and GPU in your phone affect how fast models run. Here's the hierarchy. From TokForge 1.3.8 (beta), each phone also checks and times its own CPU and GPU routes, so treat these tiers as starting points (see how your phone picks the route).

Snapdragon 8 Gen 3 and newer Snapdragon flagships

Top tier. OpenCL GPU acceleration, plus the NPU (QNN) for image generation.

Recent flagships with other GPUs (such as the Xiaomi 14T Pro and OnePlus Ace 5 Ultra)

Excellent. GPUs with Vulkan support. 8B models are usable on phones with enough RAM.

Snapdragon flagships from 2023

Very good. Solid GPU performance. Suitable for models up to 8B with good thermal management.

Upper mid-range chips

Good. Supports OpenCL. Works well with 2 to 4B models.

Mid-range Snapdragon chips

Decent. CPU + OpenCL support. Better suited for small models (0.8B to 2B).

Older or Budget Chips

CPU-only execution. Still usable with tiny models (around 1B), but slower. GPU acceleration is not available. Chips without the ARM dot-product instructions (many older budget processors) run GGUF models only; MNN models will not load on them.

How to Check Your Phone's Specs

Don't know your phone's RAM or chipset? It's easy to find:

RAM and Processor

1. Open Settings
2. Tap About Phone
3. Look for RAM and Processor (or Chipset)

Detailed CPU/GPU Info

For more detail, download an app like CPU-Z from the Play Store. It shows your exact processor model, GPU, and thermal information.

In TokForge

Once you install TokForge, the Model Manager marks which models fit your phone's RAM, and AutoForge tests settings on your phone to find the fastest, most stable ones.

Model Browser shows your RAM and ranks models by compatibility

Tablets Work Great

Thinking about running AI on a tablet instead? Tablets often have larger RAM pools and superior thermal management, making them excellent for longer AI sessions and larger models. If you have a tablet with 8GB+ RAM, it's a fantastic choice for local AI work.

What If My Phone Is Too Weak?

Don't worry. If your phone can't run the models you want locally, you have another option: use the Remote API backend. Connect TokForge to a server running the AI model, and use TokForge as the chat interface. Your phone becomes the client, and the heavy lifting happens remotely. Your messages go to that server, so use one you trust.

Real Device Examples

Want to see what your device (or a device you're thinking of buying) can actually do? Here are some typical examples:

A benchmark share card records the device, SoC, model, backend and config alongside the result. Live results from real handsets are on the phone-speed leaderboard.

Quick Answers

Can my phone run it?

If your phone has 4GB of RAM or more and runs Android 8.0 or newer on a 64-bit ARM processor, yes, at least at the small end: 4GB handles very basic models, 6GB runs 2 to 3B models and 8GB runs 4B models. The RAM tier guide above maps your exact device to what will fit.

Can I run multiple models at once?

You can have multiple models installed on your phone, but only one can run at a time. Switch between them in the Model Manager; loading a model takes a moment.

Does running AI drain my battery?

Running inference uses CPU/GPU, which does consume battery. We have not published a battery-drain measurement, so we will not put a number here. What is true: inference is a sustained CPU or GPU load, so it draws meaningfully more power than reading or scrolling, and a long session will be noticeable. Airplane mode can help a little by turning the radios off.

Will my phone overheat?

Phones are designed to handle sustained loads, and Android slows the processor down by itself if the phone gets too hot, so generation gets slower rather than unsafe. AutoForge results note when a run was thermally throttled. Tablets handle heat better due to larger surface area.

What if I want to run really large models (27B+)?

That's what the Remote API is for. Connect to a server or PC running the model, and use your phone as the interface. TokForge handles both local and remote setups smoothly.