Run KoboldCpp on an Android phone with Termux
There is no official KoboldCpp app or APK, and any APK claiming to be one is not from the project. Install Termux from F-Droid, run the official android_install.sh script, and let it download a small model; KoboldCpp then runs on the phone's CPU and you chat in the phone's browser. For bigger models, run KoboldCpp on your PC and open it from the phone instead. We tested the installer in a Termux container but not on a real phone.
Checked against KoboldCpp 1.124.2 · updated 13 Nov 2026
- gha-u24: Ubuntu 24.04 test machine · 4 vCPU · 16 GB · no GPU
On this page
There is no KoboldCpp Android APK (and why that matters)
The official KoboldCpp release has seven files, for Windows, Linux and macOS, and none for Android or ARM Linux F01. The project’s route for Android is Termux plus its install script. KoboldAI Lite also works in any phone browser at lite.koboldai.net with the AI Horde, if you only want to try it (KoboldAI, AI Horde).
Two ways to use KoboldCpp on a phone
| Criterion | On the phone (Termux) | On your PC, opened from the phone |
|---|---|---|
| Model size | Small: the script's default is a 1B model | Whatever your PC fits ([which model](/models/)) |
| Speed | Phone CPU; not measured by us untested | Your PC's speed |
| Setup | Install Termux and run a script that builds KoboldCpp (11 minutes in our container) F45 | Network address and firewall ([remote access](/remote-access/)) |
| Privacy | Local to the phone | Local to your network |
Step 1: Install Termux from F-Droid
Install Termux from F-Droid. The Play Store build of Termux is reported to be old and no longer updated, which is why the project’s guidance points to F-Droid; we did not install Termux on a phone, so we have no screenshots of this step.
Step 2: Run the official install script
The script is android_install.sh in the project’s repository. It refuses to run anywhere but Android, and it installs wget, git and python with pkg, clones the repository and builds with make -j 2 F93. We ran it with option 2 in an official Termux container: it finished with “Your KoboldCPP Installation is Complete!” in 11 minutes and built two libraries of 3.9 and 30.8 MB F45.
curl -sSL https://raw.githubusercontent.com/LostRuins/koboldcpp/concedo/android_install.sh -o android_install.sh
bash android_install.shSaving the file first matters: see the box below.
| 1 | Install and launch with the default model, Gemma 3 1B |
| 2 | Install without a model; download one later |
| 3 | Download a GGUF from a web link (needs an existing install) |
| 4 | Load a GGUF that is already in the folder (needs an existing install) |
| 5 | Rebuild an existing install |
| 6 | Exit |
Choice 1 downloads gemma-3-1b-it-Q4_K_M.gguf, asks for a context size (default 8192, accepted range 256 to 524288) and starts python koboldcpp.py --contextsize N --model <link> F93.
Step 3: Chat in the phone’s browser
Open http://localhost:5001 in the phone’s browser. KoboldAI Lite shows a mode picker the first time; pick one F31. We could not test this on a phone.
Which models run on a phone
We did not run any model on a phone, so we give no speed table. Our container could not start the program at all: it failed with dlopen failed: cannot locate symbol "__android_log_write" because Docker lacks Android’s system log library, so a real Android system is needed to run it F45. The script’s own default is a 1B model at Q4_K_M. Users report 3 to 4B Q4 models on 8 GB phones, and Q4_0 running faster on ARM; treat those as reports. Phone GPUs are not used by this route: Vulkan in Termux is experimental and users report crashes (issues 1564 and 2592). See which model and quantization.
Start it again, update, remove
Start again with python koboldcpp.py --model <file> in the folder the script created, or run the script and choose 4. To update, run it and choose 5 (rebuild); what the rebuild pulls we did not check. To remove everything, delete the folder, or uninstall Termux.
If it fails
Illegal instruction: users report it on older ARM chips (issue 624). Not reproduced by us.- Build errors: check free storage and that the install finished.
- Termux closes in the background: use the wake lock.
Error: This script is only intended for Termux on Android!: you ran it somewhere else.
For a bigger model, use your PC and open it from the phone: remote access.
Frequently asked questions
Is there a KoboldCpp APK?
No. The release has seven files for Windows, Linux and macOS and none for Android. An APK with the name is not from the project: do not install it.
Can KoboldCpp use my phone's GPU?
The Termux route runs on the CPU. Vulkan in Termux is experimental and users report crashes; we did not test it.
Which phone can run KoboldCpp?
We did not run it on a phone. The script's default is a 1B model, which suggests a phone with a few GB of free memory; users report 3 to 4B models on 8 GB phones. Treat that as a report, not a result.
Can I use SillyTavern on Android with KoboldCpp?
Yes, two ways: run both in Termux, or run KoboldCpp on a PC and open SillyTavern or Lite on the phone. We tested neither on a phone.
How much storage does KoboldCpp need on Android?
The build is small (about 35 MB of libraries in our container run), plus the repository, the build tools and the model; the default model is a 1B Q4_K_M GGUF. We did not measure the total on a phone.