No model loaded

Pick a model on the left, load it, then talk to it here

Tweak reply settings whenever you like — changes apply starting with your very next message.

Reply settings

Applied on your next message. The model stays loaded.
0 = deterministic. Past ~1.6, most small models stop forming real sentences — that's the "unhinged" zone. If output ever locks the tab up, use Stop; it's a symptom of the model never emitting an end-of-text token, not a WebGPU crash.
Limits each next word to the smallest set of candidates whose probabilities add up to this fraction, then picks among just those — so 0.95 considers almost every plausible word, while 0.5 restricts it to only the most likely ones. Lower values make output safer and more repetitive; 1.0 disables the limit entirely, leaving temperature as the only thing controlling randomness.
Above ~1.3 it starts avoiding common words entirely, which reads as "unhinged" for a different reason than temperature does.