SGLang

SGLang 0.5.22 speeds up DeepSeek-V4.1 serving by up to 2.3x on H200, adds a 10x faster model loader and runs on Apple Silicon

SGLang (GitHub) published this on Oct 18, 2026 (source: github.com).

SGLang v0.5.22 bundles 950 pull requests from 245 contributors. The team says DeepSeek-V4.1 gets up to 2.3x more throughput per GPU on H200 and now runs on AMD MI350X and MI355X, a new InstantTensor loader loads models 10x faster, and PD disaggregation gets up to 27% more throughput with half the p95 time to first token on long prompts. New model support includes Cloudflare's Clef, Perplexity's PPLX-Decider, EmbeddingGemma 2, HunyuanImage-3.0-Instruct and Kandinsky 6, and SGLang now also runs on Apple Silicon Macs with the standard PyTorch backend. The release fixes a bug where hybrid linear models could resume from the wrong state when speculative decoding and structured output were used together, and now requires Python 3.11 or newer.