miaai-lab/deepseek-v4-flash-dspark-2x-dgx-spark is a recipe repository for running DeepSeek-v4-Flash 0731 on two DGX Spark units. Its README presents DeepSeek-V4-Flash-Vision-Exp as the run target and explains a vLLM TP=2 configuration. It is not a release announcing the changes of a new version; the core is the run method and preparation procedure documented in the repository README.
This configuration is worth consulting when you want to run a large model on local hardware by leveraging DSpark quantization. The README lists the speculative decoding path, the 1M token context limit, and the image input path. It also guides you through preparing the network and image of the two nodes and configuring the model cache before an actual run. Since it covers not only the run commands but also the status-checking process, you can follow the whole flow from preparing the two-node setup to bringing it up.
That said, the README explicitly states that the official weights contain no video encoder. The image input path and the absence of a video encoder must be looked at separately, and this fact alone does not mean video input support can be expected. So this repository is best treated as reference material for confirming the run configuration and image path described in the documentation.