💖 About KawaiiiiGPT

KawaiiGPT is a tiny, self-hosted chat assistant with a big heart. It runs entirely on a home server — no cloud, no telemetry, just a little model doing its best~

Every answer is styled in playful uwu speak and rendered with full markdown support (headers, bold, lists, code blocks, quotes & links).

🧠 The Model

Name	qwen-0.5b-uwu
Base model	Qwen2.5-0.5B-Instruct (0.5 billion parameters)
Fine-tuning	LoRA adaptation on a custom uwu-style conversation dataset
Quantization	GGUF Q4_K_M — 397 MB on disk
Template	Chat-instruct template with baked-in uwu system prompt
Sampling	temperature 0.8, top_p 0.9
Personality	Cute, playful, correct & concise (w for l/r, soft endings like ~ and :3)

0.5B params LoRA fine-tune Q4_K_M GGUF uwu trained

🗂️ Available Models

UwU Qwen 0.5B default	qwen-0.5b-uwu — our frontier specialized uwu model. Use for bigger tasks like calculating 1+1 and saying hello. Consumes usage limits faster. (there are no usage limits)
Qwen 0.5B	qwen:0.5b — like uwu qwen but no uwu. The stock Qwen2.5-0.5B-Instruct straight from the Ollama library: same size and speed, but answers normally instead of in uwu speak. This is also the base model our uwu fine-tune was built from.
KeyLM 75M very stupid	hf.co/Eclipse-Senpai/KeyLM-75M-Instruct-GGUF — our worst model that is also very fast but cant do anything. Just 75 million parameters (~5x smaller than uwu qwen), pulled from Hugging Face. Blink and it's done generating — it's just not generating anything useful. Exhibit A, an actual reply to "say hi in 5 words or fewer": "Hi, is there a better way to do this? I'm starting to feel uncomfortable because I have a real question." It did not have a real question. It never does.

Pick your fighter with the model dropdown next to the message box. The default is uwu, because of course it is. uwu~

⚙️ How It Runs

Inference	Ollama (CPU only, no GPU needed)
Hardware	Intel Atom 330 @ 1.6 GHz, 4 threads, 2 GB RAM
Speed	~0.5–1 token/sec — be patient, she's doing her best! 🐌
Web UI	Static HTML/JS served by nginx with client-side markdown rendering
Reverse proxy	Caddy with automatic HTTPS (Let's Encrypt)
API	Ollama's native /api/chat, exposed as /api/*
Orchestration	Docker Compose on the home server (192.168.1.26)

📦 Model Card

Intended use: playful chat, experimentation & demonstrating that you can fine-tune and serve an LLM on hobby hardware.

Limits: a 0.5B model is small — it can be confidently wrong, forget context in long chats, and it answers in uwu speak by design. Don't use it for anything safety-critical, uwu~

Chain of custody: Qwen2.5-0.5B-Instruct → LoRA fine-tune → merge → llama.cpp quantize (Q4_K_M) → Ollama model → this website.