Releases·Tool Release
SGLang v0.5.21 ships 779 PRs, PD hot-switching, Rust prefix cache, and DeepSeek-V4.1 support
The SGLang project shipped v0.5.21, adding support for new LLM, VLM, and diffusion models including DeepSeek-V4.1 Flash, GigaChat 3.5, and DiffusionGemma. Key features include on-the-fly prefill/decode switching, a Rust-based prefix cache by default, new Decisions and Score APIs, and throughput gains for DeepSeek and Kimi models.
0 engagement1 sourceFri, Oct 2, 2026, 01:09 AM
AI Overview
- SGLang v0.5.21 merges 779 PRs from 227 contributors (sgl-project, 2026-10-02).
- Adds DeepSeek-V4.1 Flash, GigaChat 3.5, MiMo-V2.6, Ling-3.0-flash-VL, and diffusion models like DiffusionGemma and Qwen-Image 2.1.
- PD instances can switch between prefill and decode on the fly without restart (PR #28403).
- Prefix cache now runs on a Rust core by default (PR #39627).
- DeepSeek-V4.1 gets 22% faster first token on long prompts; Kimi K3 gets 20.6% higher prefill throughput in PD serving.
- New Decisions API (/v1/decisions) turns an LLM/VLM into a low-latency classifier and scorer; Score API scores all candidates in one request.
Related
Releases·Tool Release
SGLang v0.5.17 adds day-0 support for Kimi K3, a 2.8T-parameter LatentMoE model
SGLang v0.5.17, released August 8, 2026, includes day-0 support for Kimi K3, a 2.8T-parameter multimodal LatentMoE model with 896 experts and 1M-token context. The release integrates advanced serving features like DCP, DSpark speculative decoding, and KDA-aware prefix caching, verified on NVIDIA GB300 and AMD MI35x.
SGLang v0.5.17, released August 8, 2026, includes day-0 support for Kimi K3, a 2.8T-parameter multimodal LatentMoE model with 896 experts and 1M-token context. The release integrates advanced serving features like DCP, DSpark speculative decoding, and KDA-aware prefix caching, verified on NVIDIA GB300 and AMD MI35x.
0 engagement1 sourcegithubSat, Aug 8, 2026, 12:19 AM
Releases·Tool Release
SGLang v0.5.19 ships beam search, DeepEP v2, and Qwen3.8 support across 786 PRs
The SGLang project released v0.5.19, a large community-driven update adding support for new models including Qwen3.8 (2.4T-A95B) and Qwen3.8-27B. The release aggregates 786 pull requests from 214 contributors and points users to the project cookbook for the full supported-model list.
The SGLang project released v0.5.19, a large community-driven update adding support for new models including Qwen3.8 (2.4T-A95B) and Qwen3.8-27B. The release aggregates 786 pull requests from 214 contributors and points users to the project cookbook for the full supported-model list.
0 engagement1 sourcegithubSat, Sep 5, 2026, 02:27 AM
Releases·Tool Release
SGLang v0.5.18 adds Muse Glimmer, Intern-S2-Mobius, SANA-Video support
SGLang v0.5.18 is released with support for new models including Muse Glimmer, Intern-S2-Mobius, and SANA-Video. The release includes 710 PRs from 212 contributors, expanding the inference engine's model coverage.
SGLang v0.5.18 is released with support for new models including Muse Glimmer, Intern-S2-Mobius, and SANA-Video. The release includes 710 PRs from 212 contributors, expanding the inference engine's model coverage.
0 engagement1 sourcegithubSat, Aug 22, 2026, 12:09 AM
Releases·Tool Release
vLLM v0.30.0 ships DeepSeek-V4.1-Flash support, weight-cache fast start, and watermarking in 762-commit release
The vLLM project released v0.30.0, featuring 762 commits from 315 contributors (104 new). The release adds support for DeepSeek-V4.1-Flash, GLM-5.3-Flash, Kimi K3, Qwen3.8-Flash-Next, and Cohere Compass, plus a persistent per-GPU weight-cache daemon for fast engine restarts, Gumbel-max watermarking, and the HiSparse host-resident KV tier.
The vLLM project released v0.30.0, featuring 762 commits from 315 contributors (104 new). The release adds support for DeepSeek-V4.1-Flash, GLM-5.3-Flash, Kimi K3, Qwen3.8-Flash-Next, and Cohere Compass, plus a persistent per-GPU weight-cache daemon for fast engine restarts, Gumbel-max watermarking, and the HiSparse host-resident KV tier.
0 engagement1 sourcegithubTue, Sep 22, 2026, 05:20 AM
Releases·Tool Release
vLLM v0.27.0 adds Kimi K3 support with full-stack integration
vLLM v0.27.0 ships with 561 commits from 242 contributors, adding full-stack support for Kimi K3 including kernels, frontends, and quantization. It also introduces Qwen3.5 models and K-EXAONE-2.0-750B, expanding the inference engine's model coverage.
vLLM v0.27.0 ships with 561 commits from 242 contributors, adding full-stack support for Kimi K3 including kernels, frontends, and quantization. It also introduces Qwen3.5 models and K-EXAONE-2.0-750B, expanding the inference engine's model coverage.
8 engagement1 sourcegithubMon, Aug 10, 2026, 09:18 PM