// Android Device Check

Can your Android phone
run AI?

Samsung, Pixel, OnePlus, Motorola — the answer depends on the chipset, not the brand. Snapdragon, Tensor, Dimensity and Exynos behave very differently under a sustained AI load. Pick your phone for an honest breakdown.

> On Android, the SoC and thermals decide everything — two phones with the same RAM land in different tiers if one throttles and the other doesn't.
iPhone / iPad Mac Linux Android

Not sure which chip? Settings > About phone > Processor, or check the spec sheet for your exact model.

Settings > About phone > RAM, or check your model's spec sheet. Ignore "RAM Plus" / virtual RAM — swap to UFS storage is far too slow for live inference.

info On Android, tiers are driven by chipset and sustained thermals first, RAM second. A flagship SoC that throttles after 60 seconds can land below a cooler mid-range chip for long sessions — and the manufacturer's firmware decides how aggressively that throttling and background-killing happens.

// Before you start

What nobody tells you about Android local AI

Android isn't one platform — it's hundreds of hardware-and-firmware combinations. These six factors decide whether on-device AI is smooth or painful.

// Manufacturer / OEM
The same chip behaves differently per brand
An identical Snapdragon performs differently on a Samsung, OnePlus or Motorola. Each OEM ships its own thermal profile, RAM management and power policy. The brand on the box changes the AI experience as much as the silicon inside it.
→ Judge by SoC + OEM, not RAM alone
// Thermal throttling
Sustained inference cooks the SoC
Local AI is a sustained max-load workload. Phones have almost no cooling, so the SoC throttles within seconds to minutes. 7-series and Tensor G2 chips throttle hardest; only the best-cooled 8 Gen 3/4 flagships hold clocks under long sessions.
→ Expect speed to drop after the first minute
// Background limits
Android kills your inference task
Android aggressively kills background processes to save memory and battery. A model server in Termux can be killed the moment you switch apps. Use termux:boot + a wake-lock to keep it alive, or your API silently dies mid-request.
→ termux-wake-lock + termux:boot to survive
// Firmware variation
Same model, different behavior
OneUI, OxygenOS, Pixel stock and Motorola's build each handle memory, NPU access and background tasks differently — and behavior shifts between firmware updates. There's no single "Android" baseline to rely on the way there is on iOS.
→ Test on your exact build; don't assume
// Storage (UFS 3.1)
Storage speed gates model loading
Loading a multi-GB model is storage-bound. UFS 3.1 / 4.0 flagships load a 4 GB model in seconds; older UFS 2.x or eMMC budget phones take much longer and stutter. Slow storage also makes any "virtual RAM" swap useless for inference.
→ UFS 3.1+ for usable load times
// Power draw
Heavy use kills the battery
Running a local LLM or hosting an API drains the battery fast — roughly 2–6 hours of intensive use, and it generates real heat. Long sessions realistically need to be plugged in, which itself adds heat and can deepen throttling.
→ Plan to run plugged in for long jobs

// How to run it

The Android local-AI toolchain

Termux
All phones · Linux environment on Android
The foundation. A full terminal and package manager that runs Python, FastAPI, and local LLMs without root. Almost every on-device AI workflow on Android starts here. Install from F-Droid (the Play Store build is outdated).
pkg install python clang cmake git
llama.cpp
All phones · ARM64 CPU inference
The universal engine. Runs GGUF models on the ARM CPU with NEON acceleration. Always download Q4_K_M or Q8_0 quantized files — FP16 is roughly double the memory and won't fit. Pin threads to the big cores for best throughput before throttling kicks in.
./llama-cli -m model.gguf -t 6 -p "Hello"
Ollama (in Termux)
All phones · Easy wrapper for llama.cpp
Manages downloads and exposes a local REST API automatically. Slightly slower than hand-tuned llama.cpp but far simpler. Pair with termux-wake-lock so Android doesn't kill the server when you leave the app.
termux-wake-lock && ollama serve
ngrok / VPS
Optional · Exposing your local API publicly
To reach your phone's API from elsewhere, tunnel it with ngrok. For anything 24/7, a cheap VPS (~$5/month) is more reliable than fighting Android's battery and background-task limits on the phone itself.
ngrok http 11434

// The four tiers

What your phone unlocks

Budget
Snapdragon 4-series · Dimensity 6/7000 · Exynos 850 · Helio · <6 GB RAM
Proof-of-concept, not a daily driver. The smallest models (Qwen3 0.6B, Llama 3.2 1B) run slowly via Termux + llama.cpp. Budget SoCs throttle hard and lack the memory bandwidth for anything above ~3B. Educational value high, real productivity low.
Qwen3 0.6B · 0.4 GB Llama 3.2 1B · 0.7 GB Llama 3.2 3B ⚠ slow
Mid-Range
Snapdragon 7-series · Dimensity 8000 · Exynos 1xxx · 6–8 GB RAM
A usable everyday tier with patience. 3B models run interactively when cool; 7B models load on 8 GB but throttle under sustained load. Thermal management is the limiting factor here — a 5-minute session is a very different experience from a 30-second one.
Llama 3.2 3B · 2 GB ★ Phi-4 mini · 2.2 GB ★ 7B ⚠ throttles on sustained use
Flagship
Snapdragon 8 Gen 1/2 · 8+ Gen 1 · Dimensity 9000 · Tensor G2/G3 · 8–12 GB
The sweet spot for Android local AI. 7B models run at genuinely interactive speed with INT4/INT8 quantization. Tensor chips are tuned for ML but throttle sooner than the best Snapdragons. With 12 GB and good cooling, short bursts of 13B are within reach.
Qwen 2.5 7B · 4.8 GB ★ Mistral 7B · 4.5 GB ★ Phi-4 mini · 2.2 GB 13B ⚠ tight, throttles
Flagship Pro
Snapdragon 8 Gen 3 / 8 Elite · Dimensity 9300/9400 · Tensor G4 · 12–16 GB, UFS 4.0
The ceiling of what Android runs today. Best-in-class NPUs and the strongest sustained thermals handle 13B–14B models with good cooling, and 7B runs comfortably for long sessions. UFS 4.0 storage means multi-GB models load in seconds. Still battery- and heat-bound for marathon use.
Qwen 2.5 14B · ~9 GB ★ Llama 3.1 8B · 4.7 GB ★ INT4/INT8 NPU-accelerated Plugged in for long jobs
lock

Local AI on Android is private by default. llama.cpp and Ollama in Termux run entirely offline — no API key, no account, no telemetry. Your prompts stay on your phone.