SGLang v0.5.22: DeepSeek-V4.1 throughput up 2.3x, model loading 10x faster

The open-source large language model inference framework SGLang released version v0.5.22 on October 19, 2026. This update focuses on optimizing inference throughput and loading speeds for mainstream open-source models, while expanding native support for AMD and Apple hardware.

In terms of performance, SGLang reports that DeepSeek-V4.1 achieves up to 2.3x higher throughput per GPU on NVIDIA H200. The release also introduces the InstantTensor model loader; official tests show that under a storage bandwidth of 50 GB/s, the loading time for DeepSeek-R1 on 8x H200 GPUs dropped from 160 seconds to 15.3 seconds, representing an approximately 10x speedup. Furthermore, optimizations to the Prefill-Decode (PD) disaggregation architecture yield up to 27% higher throughput and halve the p95 Time To First Token (TTFT) on long-prompt workloads.

Regarding hardware ecosystems, SGLang now supports running on Apple Silicon Macs using the standard PyTorch backend alongside MLX. This marks the extension of high-performance inference capabilities into consumer desktop devices. Additionally, DeepSeek-V4.1 is now supported on AMD MI350X and MI355X GPUs, where SGLang claims DSpark decoding delivers up to 1.9x the throughput of plain decode.

It should be noted that the reported improvements in throughput and loading speed are based on self-tested benchmarks by the SGLang development team under specific hardware configurations, and have not yet been independently reproduced or verified by third parties.