What is qwen3.6-35b-a3b-ud-q8_k_xl_dgx-spark-recipe?

This repository is a small set of shell scripts that make it easier to run a specific large AI language model on your own hardware, rather than through a cloud API. The model is Qwen3.6-35B-A3B, a large open weight model released in a compressed file format called GGUF, and the scripts start and stop a program called llama-server, which is the piece of software that actually loads the model and serves it. The scripts are built around powerful workstation style machines with 96 to 128 gigabytes of graphics card memory, such as NVIDIA's DGX Spark, since a model this size needs a large amount of memory just to load, plus more for its very long 256,000 token context window and its vision capabilities. The start script automatically figures out where the llama-server program is installed, checks whether the roughly 40 gigabyte model file and its smaller companion vision file are already present, and downloads whichever ones are missing directly from Hugging Face, resuming automatically if a download is interrupted. It also checks that the server has actually become healthy before reporting it is ready, avoids starting a second copy if one is already running, and keeps running in the background even if you close the terminal window. A matching stop script shuts the server down cleanly. Important to note: this project does not include the llama-server program itself, or the AI model weights. Both need to be obtained separately, with llama.cpp built from source with CUDA support for your specific graphics card, and the model files downloaded either automatically by the start script or manually ahead of time. A Hugging Face account token is recommended to avoid download rate limits. Once running, the server exposes an API compatible with OpenAI's chat format, so it can be used as a drop in replacement for tools that expect an OpenAI style endpoint, letting you send chat messages and get responses from your own locally running model.