deepseek-ai/DeepSeek-V3.2-Exp ? reverse-engineered prompt

Build me a local demo for DeepSeek V3.2 Exp that lets me run the model from my own downloaded weights and chat with it in a simple terminal interface.

I want a setup that can first convert Hugging Face weights into the format your demo needs, then launch an interactive generation script so I can talk to the model and see responses right away. Please make it clear how to choose the number of GPUs or model parallel workers, and keep the instructions easy to follow for someone who just wants to get it running.

If there are any tricky model details that matter for correctness, like the attention or position embedding setup, please handle those properly in the demo code. Also include a clean way to configure the model and make sure the code is ready for long context use. If you need to check the latest docs or examples for the supported serving options, go ahead and look them up online.