How I created a private, self-hosted ChatGPT on the NVIDIA DGX Spark: Qwen3.6 27B, long-term memory, and cited web search.
For years I have carried around a quiet wish: one day I want an AI of my own, running on hardware I can touch, holding conversations that live in a file I can back up myself. When NVIDIA shipped me a DGX Spark, the first thing I did was make that wish concrete and build my own ChatGPT.
If you cant read the article because of paywall then please click here
The Spark itself deserves a short introduction: it is a small gold box built around the GB10 Grace Blackwell superchip with 128 GB of unified memory, which means the CPU and GPU share one pool and a 17 GB model leaves enormous headroom for context. Mine sits headless under my desk with no monitor and no keyboard attached, I reach it over SSH from my laptop, and Ollama installed on it without any drama.
For the model I went with Qwen3.6 27B, the dense 27B that the Qwen team released in April under Apache 2.0. Two of its properties ended up shaping the app: it accepts image input, which makes attachments in the chat worth having, and it exposes a real thinking toggle, so sending think…
