tonyd2wild/deepseek-v4-flash-dspark-1m-nvfp4-kv-2x-dgx-spark

This project is a detailed setup recipe, not a piece of software you install with one command. It explains, step by step, how to run a large AI language model called DeepSeek V4 Flash across two specialized computers known as DGX Spark machines, so it can handle extremely long conversations, up to about one million tokens of text, while still responding quickly. The two machines work together as a single system using a technique called tensor parallelism, splitting the model's calculations between them. The recipe also uses a memory saving format for the model's working memory, called the KV cache, so long conversations do not use up all the available graphics card memory. On the hardware described, a single conversation can be answered at roughly 50 to 60 words per second, and the recipe includes benchmarks showing it stays stable even when many conversations run at once, without the model repeating itself or switching into the wrong language. The README is written for people running agent systems, meaning automated AI assistants that call this model as a backend, on a two machine Blackwell class DGX Spark cluster. It assumes you already have that specific hardware, a fast network connection between the two machines, and familiarity with Docker containers. Setup involves editing a configuration file with your machines' network addresses, then running a series of scripts in order: one to build the software container, one to download and check the AI model files, and one to start the service on both machines at once, with the second machine starting first to avoid a startup race condition. Once running, the model is reachable over the network like a regular web API, and the README includes commands to confirm it started correctly and is using the expected amount of memory. The project also includes a separate troubleshooting guide for cases where the model repeats text, drifts into Chinese output, or leaks internal formatting into its replies, since earlier versions of this setup ran into those problems. The full README is longer than what was shown.