Finally found a way to stop sending everything to OpenAI for privacy reasons. Using that DGX Spark setup with vLLM and LiteLLM sounds like a dream for keeping data in-house. Seeing 359 tokens per second across eight users using Qwen3.6 is wild, especially if you're just running it through the standard n8n nodes. I wonder if the NVFP4 format is going to be a pain to set up on older hardware though. Might try playing around with some local MoE models later tonight if i can get my memory bandwidth figured out. It's much better than worrying about cloud API terms changing on you mid-workflow.
more here:
https://blog.n8n.io/n8n-nvidia-dgx-spark/