SGLang 0.5.22: Apple Silicon via PyTorch, Python 3.11 now required

SGLang v0.5.22 was published on GitHub on October 19, 2026 (UTC). The notes credit 950 PRs from 245 contributors. SGLang now runs on Apple Silicon Macs with the standard PyTorch backend, alongside MLX. The full notes describe this as an Apple Silicon (MPS) platform added to the standard Torch model runner (#36780). The release requires Python 3.11 or newer. On Python 3.10, `pip install sglang` does not fail. It resolves to the last release that supported 3.10 (#43021). The notes say models load 10x faster with a new InstantTensor loader, an optional extra (`sglang[instanttensor]`). The only measurement they give is DeepSeek-R1 on 8x H200, loading in 15.3 s instead of 160 s from 50 GB/s storage (#40453). The highlights say MiniMax-H3 and Qwen-Image 2.1 run in ComfyUI with SGLang Diffusion, which also loads community GGUF checkpoints and LoRAs. MiniMax-H3 in ComfyUI was already listed in v0.5.21. New models include Clef / Clef Flash, PPLX-Decider v1 / v1.1, EmbeddingGemma 2 and LLaDA2.2. New diffusion models include HunyuanImage-3.0-Instruct, Kandinsky 6, Wan-Animate-2, SANA-Video 2.0 and LLaDA-Image. The notes list one important fix: hybrid linear models could resume from wrong mamba states when speculative decoding and structured output were used together (#43029). Where the highlights name hardware for speed gains, it is H200, B200, B300 and AMD MI350X / MI355X datacenter GPUs. We have not benchmarked this release. The previous release, v0.5.21 (October 2, 2026), added a Decisions API (`/v1/decisions`). It also added setwise scoring to the existing Score API (`/v1/score`), so one request scores all candidates, and made a Rust core the default for the prefix cache.