raullenchai/twinspark is a recipe repository for running DeepSeek-V4-Flash-0731 (284B MoE) on two NVIDIA DGX Sparks. It documents how to connect the two machines into a self-healing two-node vLLM cluster. The repository is written in Shell and is MIT licensed. The latest release version is not stated in FACTS.
The README describes a setup that connects the two DGX Sparks with RoCE and runs the model with vLLM tensor parallelism. It also includes a configuration confirmed to recover after a reboot. According to the author's measurements with DSpark speculative decoding applied, single-stream generation runs at about 75 tok/s and prefill throughput at about 1.7k tok/s. It also provides a use case for connecting Codex CLI to the local endpoint, a network pre-check procedure, and notes on separate benchmarking.
The scope covered by this recipe is the combination of DeepSeek-V4-Flash-0731 and two DGX Sparks. The performance figures are the author's measurements recorded in the README, so there is no basis for generalizing them to other models or hardware. It is therefore more appropriate to view it as a reference case for large-scale MoE serving configurations.