GPUStack is an open-source manager for GPU clusters used for AI model serving. The public repository is gpustack/gpustack, and the latest release listed on GitHub is v2.2.3. However, the changes bundled into this version are not part of the material summarized here, so what actually differs must be confirmed by reading the repository's release notes directly.
The implementation language is Python and the license is Apache-2.0. What it manages spans GPU clusters in on-premise servers, Kubernetes clusters, and cloud environments, so GPU resources in several locations can be handled in one place. Its core role is orchestrating the workers and inference engines of multiple clusters to distribute model services. Inference engines can include vLLM, SGLang, TensorRT-LLM, and custom engines, and it also provides GPU instances that can be requested and accessed directly over SSH whenever needed. Automatic failover, load balancing, monitoring, authentication and access control, and an OpenAI-compatible API are all bundled in. The current GitHub star count is 5862.
That said, the provided README summary contains no performance benchmarks and no figures that show real production deployment scale. The 5862 stars and the momentum score recorded in the material (4.6 out of 10) are only signals of interest, not validated performance. No comparison target is presented either, so it is hard to tell from this material how it compares with other managers. How far a mix of engines can be stretched in an environment that mixes on-premise and cloud requires a closer look at the release notes and the source code. What can be confirmed is the breadth of what it manages and the Apache-2.0 license; performance and operational scale have to be left to a later judgment.