Karabasan - Local AI Models

Karabasan is a local AI model operator for Android. Bring your own GGUF models and run them entirely on your phone. No account, no ads, no cloud.

Karabasan does not ship any AI model of its own. You download or copy compatible model files onto your device, and the app runs them with its built-in llama.cpp engine. Nothing you type leaves your phone, so it suits private notes and anything you would rather not send to a cloud service.

What you can do

Chat: talk to any compatible GGUF chat model. Show the model's thinking steps, keep replies in one language, and watch the Ctx % badge to see how full the context window is. Conversations are saved so you can return to them.

Chat for Two: load two models side by side and send the same message to both. A quick way to compare models on your own questions.

Image: generate pictures with Stable Diffusion checkpoints in .safetensors, .gguf or .ckpt form. Stable Diffusion 1.5 is the recommended default. SD-Turbo and SDXL-Turbo trade quality for speed.

Voice: have a spoken conversation with an LFM2.5-Audio model. It understands English. Add a Whisper speech model and a chat model to talk with it in other languages.

3D: describe an object and LLaMA-Mesh writes it as an OBJ mesh. Rotate it in the live preview, then export it as .obj or as DXF for AutoCAD and other CAD tools.

File: ask questions about a PDF, Word .docx, PowerPoint .pptx or plain text file already on your device. Works fully offline and cites the pages the answer came from.

Web Search: the one opt-in exception to offline use. Connect a search provider you choose (Brave Search, SearXNG, or a custom endpoint). Only your search query is sent there. The model that reads the results and answers with sources stays on your phone.

Getting models

Pick from the in-app download list, which includes Qwen3, Llama 3, Gemma 3, Gemma 3n, Phi-4 mini, SmolLM3 and DeepSeek-R1 Distill, or paste a Hugging Face link to any single GGUF file. A comparison table on the start screen shows parameter count, quantization, file size, and RAM needs at a glance.

You are responsible for obtaining models from legitimate sources and for following each model's license and usage requirements.

Local processing

All computation happens on your device's processor. Speed depends on RAM, processor, model size and quantization. As a rough guide, 0.6B to 4B models run on almost any modern Android phone. 8B models need a high-end phone with 12 GB or more RAM.

Karabasan warns before loading a model that is likely too large for your free memory, shows live CPU, RAM, and temperature readouts, lets you cap how many cores a model may use, and frees all loaded models with one tap.

Free, private, no sign-up

Karabasan is free and shows no advertisements. No email address, password, or account is needed. The interface is available in English, Turkish, Spanish, and Brazilian Portuguese.

Please note

AI models can produce inaccurate information and can be confidently wrong about anything after their training date. Verify important information independently. Karabasan is not a substitute for a doctor, therapist, lawyer or financial advisor. Users are responsible for the models and files they load and for complying with applicable laws and model licenses.

Other apps developed by OBN Apps:

PDF Editor: https://play.google.com/store/apps/details?id=com.obn.pdf
Text Editor: https://play.google.com/store/apps/details?id=com.obn.texteditor
Delivery Text: https://play.google.com/store/apps/details?id=com.obn.deliverytext
Memory Game: https://play.google.com/store/apps/details?id=com.obn.memorygame
Timer Countdown: https://play.google.com/store/apps/details?id=com.obn.timerapp
Mpedance: https://play.google.com/store/apps/details?id=com.obn.mpedance
Camera: https://play.google.com/store/apps/details?id=com.obn.camera
Password Manager: https://play.google.com/store/apps/details?id=com.obn.password