SGLang v0.5.23 Released with Expanded Model Support and Major Performance Boosts

SGLang v0.5.23 has been released with contributions from 257 developers, introducing support for various new LLM, VLM, and diffusion models along with significant performance optimizations. Key updates include low-latency execution modes via megakernels, Paged Experts for running large MoE models on consumer GPUs, and Google TPU support. These performance enhancements and expanded hardware compatibility significantly reduce inference costs and latency for large language and multimodal models across various platforms. This makes high-performance AI serving more accessible for both enterprise deployments and hardware-constrained environments. The release adds support for models such as MiniCPM-V 4.7 and DeepSeek-V4.1 optimizations, bringing up to a 1.5x throughput increase and faster first-token latency. Users can upgrade via pip using the command `uv pip install --prerelease=allow sglang==0.5.23` or utilize pre-built Docker containers across NVIDIA, AMD, Intel, and Google TPU platforms.