FlashML-org/FreeToken ? reverse-engineered prompt

Build me a FreeToken style app that lets people run huge MoE language models locally on a gaming PC or workstation, with a simple desktop interface and a CLI. I want it to make setup easy, let me chat with models, load supported open weight models, and expose an API that works with OpenAI and Anthropic style clients so I can plug it into tools like coding agents.

It should be fast on NVIDIA RTX cards, use GPU and CPU memory intelligently, cache experts and KV state so it does less repeat work, and let VRAM shift between model weights and context without restarting. If you need to, look up the current docs online and match the install and quick start flow people would expect. Keep it practical, polished, and ready for real local use.