I keep seeing people sayi they can run Kimi K3 in personal PC so a small cluster server rented. I understand that will have a lots of performnace drop and trade-off but why need to do so if you can just spend a litte money to use full K3 in OpenRouter, I'm just curious...

Interesting approach running Kimi K3 in 29GB — curious what quantization/precision tradeoffs you had to make to get it there. I've been dealing with similar memory constraints running client-side AI models in the browser (background removal via ONNX runtime) — the tradeoff between model size and quality never really goes away, just moves where you hit it.