ChatBack AI Model Mirror

Stable mirrors of on-device AI models used by the ChatBack Android app. All inference runs 100% locally on the user's device — no data is sent to any server.

See NOTICE.md for full attribution and licensing.

Models

Base Model (Required)

Model	File	Size	License
Gemma 3 1B-IT (q4_block128, ekv4096, LiteRT)	`gemma3-1b/Gemma3-1B-IT_multi-prefill-seq_q4_block128_ekv4096.task`	~556 MB	Gemma Terms of Use

Powers all chatroom AI features: smart replies, tone warnings, draft assistant, auto-translation, summarisation, and AI Chat (when no large model is activated). True 4 096-token KV cache baked into the .task file. Fast, low-RAM (~2 GB+), GPU/NPU-accelerated via LiteRT.

AI Chat Upgrade (Optional — requires base model)

Model	File	Size	License
Gemma 4 E2B-IT (LiteRT)	`gemma4-e2b/gemma-4-E2B-it.litertlm`	~2.5 GB	Gemma Terms of Use
Gemma 4 E2B-IT Abliterated (NSFW, LiteRT)	`gemma4-e2b/Gemma-4-E2B-it-abliterated.litertlm`	~2.5 GB	Gemma Terms of Use

Best AI Chat quality. Loaded only while AI Chat is open; base model restores automatically when you leave AI Chat. Requires 3 GB+ RAM. The abliterated variant has safety alignment removed and is only visible when NSFW mode is enabled in the app.

Text-to-Speech

Model	Size	License
Piper TTS — Amy, Lessac, LibriTTS-R, Ryan, Joe (ONNX medium)	~178 MB total	MIT
Kokoro TTS English v1.0 (q8f16 ONNX, 28 voices)	~100 MB	Apache 2.0

Kokoro uses mixed int8-weight/float16-activation quantization for better audio quality vs pure int8. The app uses its own Java IPA phonemizer — no espeak-ng dependency.

Download URLs used by the app

# Base model
https://huggingface.co/Chazzo2008/chatback-ai-models/resolve/main/gemma3-1b/Gemma3-1B-IT_multi-prefill-seq_q4_block128_ekv4096.task

# Large model upgrade
https://huggingface.co/Chazzo2008/chatback-ai-models/resolve/main/gemma4-e2b/gemma-4-E2B-it.litertlm

# Large model — NSFW abliterated variant
https://huggingface.co/Chazzo2008/chatback-ai-models/resolve/main/gemma4-e2b/Gemma-4-E2B-it-abliterated.litertlm