# SGLang v0.5.18

- Product: SGLang (https://whatsnew.fyi/product/sglang)
- Vendor: LMSYS
- Date: 2026-08-22
- Version: v0.5.18
- Original notes: https://github.com/sgl-project/sglang/releases/tag/v0.5.18
- Permalink: https://whatsnew.fyi/product/sglang/releases/v0.5.18

What's New is an index, not a publisher: every entry below links to the vendor's own release notes, which are the authoritative source. Entries are labelled where they are hand-curated sample data, pre-releases, or drawn from a secondary source such as a developer blog.

Reuse: the summaries, labels and curation here are © What's New. Quote freely with attribution and a link back; wholesale republication of the corpus is not permitted — terms: https://whatsnew.fyi/terms. The vendors' own release notes remain their publishers'.

---

- **added** — Support for Muse Glimmer autoregressive multimodal model
- **added** — Support for Intern-S2-Mobius autoregressive model
- **added** — Support for SANA-Video diffusion model
- **added** — Support for LingBot-Video-MoE diffusion model
- **added** — Support for LTX-2.5 diffusion model
- **added** — Support for Cosmos3 Edge & Distilled diffusion models
- **added** — Support for LongCat-Image diffusion model
- **added** — Overlapped checkpoint staging at startup with --startup-weight-load-mode overlap flag
- **added** — FlashInfer MNNVL for pure allreduce with --enable-flashinfer-pure-allreduce flag
- **added** — Native multimodal processing for Qwen VL in Rust server
- **added** — Support for inkling dspark in speculative decoding
- **added** — Support logprobs with DSpark speculative decoding
- **added** — Support logprobs with DFlash speculative decoding
- **added** — Support MegaMoE for DSpark under dp attention
- **added** — Support DSPARK speculative decoding in unified memory
- **changed** — TP LMHead allgather and scatter now uses single all-to-all for pure-DP dp-attention
- **changed** — Triton, FlashInfer, Inductor, DeepGEMM, and CUDA driver caches consolidated under SGLANG_CACHE_DIR
- **changed** — Updated torch to 2.13.0 with triton 3.7.1
- **changed** — Updated flashinfer to 0.6.17
- **changed** — Updated sgl-kernel to 0.4.6.post1

#### Highlights

*710 PRs from 212 contributors.*

**New models in this release** (see the [cookbook](https://docs.sglang.io/cookbook) for all supported models):

| Model | Type | PRs | Cookbook |
|---|---|---|---|
| Muse Glimmer | Autoregressive (Multimodal) | [#34262](https://github.com/sgl-project/sglang/pull/34262) | [link](https://docs.sglang.io/cookbook/autoregressive/Meta/MuseGlimmer) |
| Intern-S2-Mobius | Autoregressive | [#33691](https://github.com/sgl-project/sglang/pull/33691) | [link](https://docs.sglang.io/cookbook/autoregressive/InternLM/Intern-S2-Mobius) |
| SANA-Video | Diffusion | [#32921](https://github.com/sgl-project/sglang/pull/32921) | [link](https://docs.sglang.io/cookbook/diffusion/SANA-Video/SANA-Video) |
| LingBot-Video-MoE | Diffusion | [#32341](https://github.com/sgl-project/sglang/pull/32341) | [link](https://docs.sglang.io/cookbook/diffusion/LingBot-Video/LingBot-Video-MoE) |
| LTX-2.5 | Diffusion | [#34471](https://github.com/sgl-project/sglang/pull/34471) | [link](https://docs.sglang.io/cookbook/diffusion/LTX/LTX2.5) |
| Cosmos3 Edge & Distilled | Diffusion | [#31590](https://github.com/sgl-project/sglang/pull/31590) | [link](https://docs.sglang.io/cookbook/diffusion/Cosmos/Cosmos3) |
| LongCat-Image | Diffusion | [#23274](https://github.com/sgl-project/sglang/pull/23274) | |

Plus cookbook recipes for the [Qwen3.8 family](https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.8), [Ling-3.0](https://docs.sglang.io/cookbook/autoregressive/InclusionAI/Ling-3.0-flash), [Nemotron 3.5 Lightning](https://docs.sglang.io/cookbook/autoregressive/NVIDIA/Nemotron3.5-Lightning), [Dots3-Note](https://docs.sglang.io/cookbook/autoregressive/RedNote/Dots3-Note), and DeepSeek-V4-Pro-0813 ([#34809](https://github.com/sgl-project/sglang/pull/34809)).

**Overlapped checkpoint staging at startup**: Checkpoint pages now stage from storage while CUDA graphs capture. Qwen3-32B on H100 starts **8.6-11.7% faster** than serial with prefetch, and **2.38x faster (35.6s vs 84.8s)** than the plain default. Opt in with `--startup-weight-load-mode overlap` ([#32017](https://github.com/sgl-project/sglang/pull/32017)).

**TP LMHead with All-to-All**: The TP LMHead's allgather + scatter becomes a single all-to-all for pure-DP dp-attention. On DeepSeek-V4-Pro B200 decode, LMHead time drops **320us to 169us** and TPOT improves 36.97ms to 35.67ms ([#32313](https://github.com/sgl-project/sglang/pull/32313)).

**FlashInfer MNNVL for pure allreduce**: Non-fused allreduce sites now reuse the FlashInfer MNNVL workspace instead of falling back to NCCL. DeepSeek-V4-Flash TP4 decode on Blackwell gains **up to +6.9% at small batches**. Auto-enabled for DeepSeek-V3/V3.2/V4; elsewhere `--enable-flashinfer-pure-allreduce` ([#30700](https://github.com/sgl-project/sglang/pull/30700)).

**One compiled-kernel cache directory**: Triton, FlashInfer, Inductor, DeepGEMM, and CUDA driver caches all move under `SGLANG_CACHE_DIR`. The first launch after upgrading recompiles once; see Breaking Changes ([#32434](https://github.com/sgl-project/sglang/pull/32434)).

**Dependencies**: torch 2.13.0 with triton 3.7.1 ([#28836](https://github.com/sgl-project/sglang/pull/28836)), flashinfer 0.6.17 ([#33997](https://github.com/sgl-project/sglang/pull/33997)), CuTeDSL 4.6.2, fixing an FA4 startup regression on Blackwell ([#34372](https://github.com/sgl-project/sglang/pull/34372)), DeepEP now installed from released `sgl-deep-ep` wheels ([#33932](https://github.com/sgl-project/sglang/pull/33932)), and sgl-kernel 0.4.6.post1 ([#33842](https://github.com/sgl-project/sglang/pull/33842)).

*Full release notes by category below; breaking changes and known issues are at the end.*

##### Rust Server

* [mm] rust-server: native multimodal processing for Qwen VL (integrate sglang-mm, e2e): [#32365](https://github.com/sgl-project/sglang/pull/32365)
* refactor error responses into shared utils::response helpers: [#33894](https://github.com/sg

_[Truncated at 4000 characters — full notes: https://github.com/sgl-project/sglang/releases/tag/v0.5.18]_