AI chat
NekoChat
Private AI chat that runs on your phone.
Runs GPT-2, Qwen3 and IBM Granite 4.0-H models on the phone with its own C++ engine, using Vulkan, OpenGL ES or the CPU. It needs no ML framework, you can talk to it and hear it answer, and your chats never leave the device.
- Version
- 0.2.0
- Licence
- GPL-3.0
- Runs on
- Android 5.0+
- Origin
- Original to the suite
See it in action.
Granite 4.0-H 350M on the CPU 
Qwen3-0.6B on Vulkan, streaming 
Voice input, on the phone 
Read aloud with Kokoro 
Download from Hugging Face 
Weight precision
What makes it different.
Three model families, matched token for token
GPT-2, Qwen3 and IBM Granite 4.0-H (350M and 1B) give the same tokens and the same greedy replies as Hugging Face transformers. Redo a reply or edit your message, and the work already done is reused.
Uses the GPU when it can, in FP16, FP8 or FP4
Auto mode tries Vulkan, then OpenGL ES 3.1, then the CPU, and trusts a GPU only after it matches the CPU in a self-test. Weights are quantized as the model loads: FP8 halves memory with nearly identical answers, and FP4 quarters it.
Talk to it, and hear it answer
Tap the mic and IBM Granite Speech turns what you say into text on the phone; the recording never leaves it. Kokoro-82M reads replies aloud with 28 English voices, even on de-Googled phones with no text-to-speech engine.
Bring your own model
One-tap downloads of Qwen3 and Granite from Hugging Face, any GPT-2, Qwen3 or Granite 4.0-H repository, or a folder of .safetensors or .pt files. NekoChat checks config.json before it downloads anything.
Three themes.
Rose is One Dark with pastel pink, the original look. Ocean is Solarized Dark with pastel blues. Gothic Pink is near-black plum with soft pink, checked against WCAG 2.2 AA on every surface.
Rose, Ocean and Gothic Pink 
Gothic Pink, greeting by time of day 
Ocean
Built with.
- Kotlin
- Jetpack Compose
- C++17 / JNI
- Vulkan · GLES 3.1 · NEON