SparkDuet
An operating layer for a pair of NVIDIA DGX Spark machines that lets several large models share the hardware as switchable lanes, plus on-demand model loading and QLoRA fine-tuning on the idle node.
SparkDuet treats two NVIDIA DGX Spark machines, 256 GB of combined unified GPU memory linked by a 200 Gb/s private interconnect, as a small cluster with switchable lanes rather than one big GPU permanently dedicated to a single model. A single command switches which of several resident models, DeepSeek-V4-Flash, Qwen 27B, Qwen3.8-Flash-Next and GLM-5.3-Flash, is loaded into unified memory behind one shared OpenAI-compatible endpoint, and a single command reverts back.
While one model is resident, the rest of a larger library of GGUF and merged fine-tuned models stays on NVMe and loads on demand without touching the resident model, and the node that is not currently serving can fine-tune models up to about 70B parameters with QLoRA using Unsloth. The project publishes measured throughput numbers and committed benchmark artifacts rather than only claims, and readiness is confirmed by a real model listing and completion rather than a container simply reporting itself as up.
It is open source under the MIT license.