Can GPT-OSS 20B run on ROG Phone 9 Pro?
YES — Runs great
Formula estimate
What this means
Our estimate says GPT-OSS 20B should fit comfortably on your ROG Phone 9 Pro.
Download the MXFP4 version, which is about 12.1 GB. We expect it to use about 14.8 GB of the roughly 20 GB available to a model on this phone.
At the estimated speed, a roughly 300-word answer may take about 24 seconds to finish.
This result is calculated from the phone and model specifications. It has not been measured on this exact phone and setup.
16.7tokens/s
estimated · instant
14.8GB
needed at MXFP4
MXFP4
quant checked
needs 14.8 GBusable 20 GB
Snapdragon 8 Elite24 GB RAMCPU / GPUAlso sold with 16 GB
See how fast it feels
Estimated at 16.7 tokens/s — instant. A ~300-word reply takes about 24 seconds on this phone.
Where the memory goes
| Component | Detail | GB |
|---|---|---|
| Model weights | MXFP4 GGUF (12.1 GB) + mmap overhead | 12.7 |
| KV cache | 4K context window | 1.5 |
| Runtime | llama.cpp + app overhead | 0.6 |
| Total needed | at MXFP4, 4K context | 14.8 |
| Working budget | 24 GB RAM − Android system reserve | 20 |
| Headroom | remaining inside the working budget | 5.2 |
Pick your quant
| Quant | Download | Verdict | Speed |
|---|---|---|---|
| MXFP4 BEST HERE | 12.1 GB | ✓ Runs great | ~16.7 tokens/s |
Get your first offline chat working
Recommended app: PocketPal. Follow the point-and-click steps below. The speed above is an estimate, not a measurement from this exact app and phone.
1
Install or update PocketPal from Google Play. It is free and does not require an account. Use a current version so its loader supports newer model architectures.
2
Open the exact model. In PocketPal, go to Models → + → Add from Hugging Face, then paste
ggml-org/gpt-oss-20b-GGUF.3
Choose the MXFP4 GGUF file. The download is about 12.1 GB, so use Wi-Fi and keep the app open. Choose the main GGUF weights, not a vision projector, mmproj, or other helper file.
4
Tap Download, then Load. Start with a 4K (4096-token) context. Keep the app’s default Android backend for the first run.
5
Send a simple first prompt. Try “Explain why the sky is blue in three sentences.” This page estimates about 16.7 tokens/s, but that number is not a PocketPal measurement unless it carries a ✓ Verified label.
6
Confirm it is really offline. After the first reply, turn on airplane mode and ask a second question. If it still answers, the model is running on your phone.
If it does not work
- Model not listed: update PocketPal and paste the exact repository
ggml-org/gpt-oss-20b-GGUF. - App closes while loading: close other apps, restart the phone, and try 2K context. If it still closes, choose a smaller model.
- No offline reply: confirm that the MXFP4 GGUF file is loaded in the chat rather than a remote model.
Related checks
More on ROG Phone 9 Pro
GPT-OSS 20B on other phones