sgl-project/sglang is a serving framework for large language models and multimodal models. According to the GitHub release page, v0.5.22 was released on 2026-10-19. This version covers new model support and model loading changes, disaggregated placement of prefill and decoding, and an expanded scope for sparse attention. It is an update that bundles new model support together with changes to the serving process. The provided release summary does not state which specific models were added or which part of model loading changed.
The focus of this change is not only the range of model support but also the serving configuration. Disaggregated placement of prefill and decoding and the expanded scope for sparse attention are presented together in the release summary. The summary covers them alongside new model support and model loading, but does not explain the detailed implementation of each change or the conditions under which it applies. So what this release lets us confirm is the direction and the set of changes, while the concrete scope of behavior is hard to determine from the release information alone.
The release description includes measurements related to throughput and speed improvements, but those are results presented by the distributor. This analysis did not independently verify that performance. Performance improvement should therefore be treated as a claim stated in the release description rather than as a verified result. From the provided information alone it is not possible to tell in which environment or by how much things change, so the extent of the performance improvement remains unconfirmed.