GLM-5.3-Flash-NVFP4-2x-DGX-Sparks-BIZ
A community-built serving stack that runs NVIDIA's GLM-5.3-Flash NVFP4 checkpoint across two DGX Spark-class GB10 systems through a pinned vLLM build, aimed at commercially usable, pinned and reversible deployment rather than peak throughput.
This project packages a pinned vLLM build to serve NVIDIA's GLM-5.3-Flash NVFP4 checkpoint split with tensor parallelism across two DGX Spark or compatible GB10 systems connected over a network link, with published measurements taken on two specific workstation systems. It focuses on commercially usable licensing, pinned artifacts, observable checks and reversible operation rather than maximum speed.
Two serving profiles are offered: distributed defaults that serve the weights exactly as NVIDIA distributes them, and a published option that repacks the attention projections and output head to a lower precision for faster decode at a measured quality cost, as an explicit operator opt-in. Both profiles are documented as accepted for routine single- or two-sequence use since September 2026, with the accepted scope, validation results and unverified areas all listed separately.
The serving code is Apache-2.0 licensed; the model weights are MIT licensed and downloaded separately, with each artifact keeping its own terms.