Chakuli - Offline AI Chatbot

Chakuli is a private AI chatbot that runs on your phone — not in the cloud. Download an open model once, then chat offline. Your prompts, replies, photos, and history stay on the device. No account. No chat servers watching the conversation. Use it to write, code, study, translate, or think through a hard problem when you want a capable assistant without sending the chat away. HOW IT WORKS Chakuli loads Large Language Models (LLMs) locally. After a model is on the phone, answers are generated on-device. You can travel, use airplane mode, or save mobile data and still have a chatbot in your pocket. Internet is used to download models. Everyday chat with those models does not need a network. WHAT YOU CAN DO • Offline AI chat with a local LLM • Switch models: small and fast, or larger and stronger • Deep Thinking — supported models reason step by step before they answer • See & Chat — ask about a photo (Gemma 4) • Listen & Chat — record audio to transcribe, translate, or reply (Gemma 4) • Agent Skills — on-device tools on supported models. A few optional skills use the internet only if you run them • Import more models: GGUF from Hugging Face or files already on your phone • Markdown replies, chat history on-device, and a built-in speed/quality benchmark MODELS Recommended: Google Gemma 4 (E2B / E4B) through LiteRT-LM — chat, thinking, images, audio, and tools on one file. E2B is the practical pick for many 6–8 GB phones. Also in the catalog: Qwen3 and Qwen3.5, SmolLM, IBM Granite, Microsoft Phi-4 Mini, DeepSeek R1 distill, Gemma 3, plus Google Gemini Nano (system AICore on Google’s short device list — not every flagship). TWO ENGINES, ONE APP • LiteRT-LM (.litertlm) — Google’s on-device stack. Tries GPU, and NPU on some Pixels, then falls back to CPU if a path fails. • llama.cpp GGUF — broad open-model format. Production path is ARM CPU (NEON, mobile-tuned threads). Optional Vulkan GPU is off by default and limited to GPUs we have found stable (Adreno). You do not have to pick an engine. Choose a model; Chakuli loads the right runtime. SPEED, WITHOUT THE HYPE Local AI is only as fast as the phone. Chakuli is built for modern 64-bit Android: GPU where LiteRT supports it, CPU that prefers performance cores for generation, and settings for Auto / Performance / Battery, context window, and Gemma 4 speculative decode (MTP). A mid-range or flagship phone with enough RAM is comfortable. Tiny models exist for tighter memory; 8B-class files need more RAM and will feel slower. First load of a large model takes a bit; later chats are quicker. PRIVACY For local models, inference stays on the device. Chakuli does not require an account to chat. Crash reports and analytics are off unless you turn them on. Optional Tip Jar — every feature works without paying. Tips support a solo developer. GOOD TO KNOW • Android 13+, 64-bit ARM • About 4 GB RAM minimum; 6–8 GB+ recommended for Gemma 4 E2B • Storage for the model files you download (from hundreds of MB to several GB) • Older or very low-RAM phones can run small models but generation will be slower • Unlocks and some OEM “AI” features are not the same as Gemini Nano Prompt API support Take a local LLM with you. Download Chakuli and keep the conversation on the phone.