Lissa Playground

Side-by-side comparison of GGUF quantizations of two tiny pony language models — Nano (32M) and Brat (103M), trained from scratch on MLP fandom prose and supervised on roleplay dialogue. The same prompt and the same seed go to every selected quantization, so the difference you see is exactly what compression costs. Everything runs in your browser via llama.cpp compiled to WebAssembly — no server, no API key.

Model Quantizations (pick up to 3)

system — persona (required, no empty) user — message

Sampling — applies to every selected quant (same seed ⇒ same text, by design)

temperature 0.80

top_p 0.90

top_k 40

repeat_penalty 1.10

max tokens 90

seed

Perplexity of every published quantization

Perplexity measured with llama-perplexity on the same held-out 300-token English text for every file and for the bf16 original. Below Q4 the model keeps the topic and the format but loses voice and coherence; Q2 files are published for size experiments only.

Prompt format. The model was supervised on exactly <system>{persona}<user>{message}<assistant>{reply}<eos> — tags concatenated, no spaces. Its bos_token is <eos>, so no BOS is added; the published GGUF files already carry add_bos_token = false.
First run downloads the selected files into your browser cache (20–125 MB), so give it a few seconds. Later runs load instantly. CPU-only WebAssembly — expect a few tokens per second on small laptops.
Models: Nano GGUF · Brat GGUF · Nano base · Brat base