SGLang v0.5.19: Qwen3.8 support, beam search, and DeepEP v2
On September 5, 2026, SGLang v0.5.19 was released. The release includes 786 pull requests from 214 contributors. Among the new models are Qwen3.8 (2.4T-A95B), Qwen3.8-27B, Ling-3.0-flash, Ling-3.0-tiny, Spark2.5, MiniCPM-SALA, and Granite 4.2, as well as the diffusion models LongCat-Image-Edit and Edit-Turbo.
Among the system changes: beam search has been added — the beam_width parameter in a request returns the n best sequences instead of one. It works in parallel with regular requests, but is not yet compatible with speculative decoding, disaggregation, DP attention, and HiCache. The DeepEP v2 engine with ElasticBuffer for DeepSeek-V3/V4 and Qwen3-MoE in FP8 has also been added: fixed-size buffers allow decode to work under CUDA graphs even across nodes.
What it means
For those running models on their own hardware, the release brings three practical changes. First, the list of supported models has expanded: Qwen3.8-27B has been tested on the RTX 5090, RTX PRO 6000, and DGX Spark, and MiniMax-H3 on GPUs with 24 GB of VRAM, with a separate guide for tuning on consumer cards. Second, beam search provides multiple answer options for a single request — useful for tasks where you need to compare generations rather than take the first one. Third, DeepEP v2 with fixed buffers simplifies running MoE models in FP8 across multiple nodes: decode works more stably under CUDA graphs, reducing synchronization overhead.
The limitations of beam search should be considered when planning: it does not combine with a number of optimizations that are often enabled to speed up inference. For more on supported models and new deployment guides, see the CPU3D news feed.
