text model · Qwen3 · Android
Can I run Qwen3 4B on Samsung Galaxy S25 Ultra (16GB, 1TB config only)?
Yes. Qwen3 4B runs on Samsung Galaxy S25 Ultra (16GB, 1TB config only) at Q4_K_M (~3.8 GB of ~12 GB usable).
Runs at Q4_K_M using ~3.8 GB of ~12 GB usable. You have room for FP16 for higher quality.
That figure is at a 4k context and moves about ±15% as context length changes. Samsung Galaxy S25 Ultra (16GB, 1TB config only) leaves ~8.2 GB of headroom, room to step up to FP16 for higher quality.
- Q4_K_M needed
- ~3.8 GB
- Usable on device
- ~12 GB
- Device memory
- 16 GB
- Best quant
- Q4_K_M
Which quant fits
Running cost · estimate
- Power draw
- ~8 W
- Electricity / 1M tokens
- ~$0.02
- Pays for itself after
- ~3,373M tok
At ~$0.15/kWh and the estimated ~17 tok/s, a million generated tokens costs about $0.02 in electricity. Versus a hosted API at ~$0.5 per million tokens, the ~$1,619 Samsung Galaxy S25 Ultra (16GB, 1TB config only) pays for itself after roughly 3,373 million tokens, so local hardware is mostly a fixed cost, not a per-token one. TDP is the peak draw, so this is an upper bound. Assumptions.
How to run it
On Android use PocketPal AI (Polished app, download GGUF and run offline.).
- Parameters
- 4B
- Q4_K_M size
- 2.5 GB
- Q8_0 size
- 4.28 GB
- Context
- 32k
- Ollama tag
- qwen3:4b
- Memory
- 16 GB ram
- Usable for weights
- ~12 GB
- Power draw
- ~8 W
- Best runtime
- llama.cpp (PocketPal) or MLC-LLM (Adreno GPU path)
You could also run
Run Qwen3 4B on other hardware
FAQ
Can Samsung Galaxy S25 Ultra (16GB, 1TB config only) run Qwen3 4B?
Yes. Qwen3 4B runs on Samsung Galaxy S25 Ultra (16GB, 1TB config only) at Q4_K_M (~3.8 GB of ~12 GB usable).
How much memory does Qwen3 4B need?
Samsung Galaxy S25 Ultra (16GB, 1TB config only) has room to spare. At Q4_K_M the weights are ~2.5 GB; with KV cache and runtime overhead, budget ~3.8 GB at a 4k context.
What is the best tool to run Qwen3 4B on Android?
On Android, PocketPal AI (Polished app, download GGUF and run offline.) is the go-to option. NPU acceleration is limited and chip-specific; most apps run on CPU. Expect 1B-4B class.
Embed this
A live badge for your model card or README, updated as the data is. [](https://localmodel.run/can-i-run/qwen3-4b/samsung-s25-ultra-16gb) Sources
- gadgetversus.com
- github.com/a-ghorbani
- github.com/google-ai-edge
- github.com/mlc-ai
- github.com/QwenLM
- github.com/shubham0204
- github.com/Vali-98
- gsmarena.com/samsung_galaxy_s25_ultra-12750.php
- gsmarena.com/samsung_galaxy_s25_ultra-13322.php
- huggingface.co
- layla-network.ai
- ollama.com
- sammyfans.com
- sammyguru.com
- samsung.com
Weights are measured from GGUF files; KV cache and overhead are computed, so totals can vary ~15% with context and runtime. Any tok/s is a bandwidth estimate. See methodology.