The latest release of sgl-project/sglang, based on the provided repository information, is v0.5.22. SGLang is an open-source inference framework for large language models and multimodal models. That said, this information includes neither the changes in v0.5.22 nor its release date, so what was added or fixed in this version cannot be confirmed. The repository description introduces it as a high-performance serving framework, written in Python and using the Apache-2.0 license.
According to the README, SGLang covers not only large language models but also vision-language models and diffusion models. It describes itself as optimized for agentic workloads, large-scale serving, and RL rollouts. An RL rollout is the process of generating model outputs used for reinforcement learning. A separate SGLang Diffusion engine is used for image and video generation, and it is included in the repository and in the `sglang` Python package. It lists GPU, TPU, NPU, CPU, and Apple Silicon as supported targets, and it says models can be run by installing it as a Docker image or in a Python environment.
The point that draws attention is the scope covered by a single framework that handles both model serving and generation work. The provided record lists release dates from v0.5.19 through v0.5.21, and releases came at intervals of 2~3 weeks during that period. It also mentions cases of supporting new models such as DeepSeek-V4.1 Flash on the day of the release, as well as features related to prefix caching. However, this information alone is not enough to judge the specific improvements or performance advantages of v0.5.22.