0xdfi/glm-5.2-1m-4x-dgx-spark

This project is a detailed recipe and set of measurements for running a very large AI language model, GLM-5.2, on a small cluster of four NVIDIA DGX Spark computers connected together. The model itself is huge, about 744 billion parameters, and the project shows how to run it in full, without shrinking or removing parts of it, while still allowing it to read and remember up to one million tokens of text at once, which is roughly the length of several long books. The author explains that getting a model this large to run well on hardware this size takes several tricks working together. These include a compact way of storing the model's memory of a conversation so it takes up less space, a specialized attention method that speeds up how the model processes long text, and a technique that lets the model draft several words ahead at once to speed up its replies. The project also shows a tradeoff dial: you can choose to run the model with less memory context but faster replies, or more context but slower replies, and the README gives specific numbers for each setting. All of the performance numbers in the README come from actually running the system and reading the results from the software's own logs, rather than from theoretical estimates. The author reports specific speeds, such as how many words per second the model can produce depending on the type of text involved, and compares these results to other public reports of similar hardware setups, noting that this setup avoids shrinking the model the way some other attempts did. This is a niche, highly technical project aimed at people who already own or have access to DGX Spark hardware and want to run very large AI models locally rather than through a cloud service. The README does not state a software license, so anyone interested in reusing this work should check with the repository directly before assuming what they are allowed to do with it.