kindlingai/glm-5.3-flash-exlr8-1x-spark is a repository that provides a Docker deployment configuration for running GLM-5.3-Flash with gr8 on a single DGX Spark. There is no public latest release, and the README describes a configuration that combines K2 weights with an int4 key-value cache, DFlash2 speculative decoding, and a disk-based prompt cache. The repository is licensed under MIT.
This configuration walks you through choosing between a speed-first setup and a long-context-first setup when running the model on a single device. The number of concurrent requests and the context length are tunable, and the configuration given in the README handles 4 requests at a time. The prompt cache stores context on disk, and the int4 key-value cache and DFlash2 speculative decoding are also part of the runtime setup. So the core of this repository is not an introduction to a new model or a separate library, but the configuration for deploying GLM-5.3-Flash on a single device and how to tune its capacity.
The README reports 54.2 tokens per second and requests of up to 262144 tokens in the default configuration. These are measurements from the single device listed in the README, and it explicitly notes that throughput drops in the long-context configuration. So rather than generalizing these numbers as expected performance on other hardware or settings, it is more appropriate to treat them as the author's own measurements for that deployment configuration.