SGLang v0.5.16 adds DSpark, cuts GLM-5.2 KV memory and supports Inkling on day one

BanghuaZ · x · 2026-07-25

SGLang v0.5.16 ships a new speculative decoding algorithm called DSpark and several inference/runtime optimizations.

• DSpark is reported to stay ahead of MTP across concurrency levels, with 383.7 tok/s at accept length 5 on DeepSeek-V4-Pro (TP8, B300, bs=1).
• Day-0 support is added for Inkling, the first open-weights model from Thinking Machines Lab, with verified recipes for NVIDIA Blackwell, H200, and AMD MI350X/MI355X.
• Other newly supported models include LongCat 2.0 FP8, JetBrains Mellum v2, Pi0.5, and LongLive 2.0.
• The release also introduces one shared prefix cache for SWA/Mamba/DSA models, GPU cache splitting for GLM-5.2, lighter speculative decoding for GDN, and the first correct KDA MTP path on Blackwell.

Original post →

More from Infra

• SpaceX strikes $40B deal with Apollo Global to lease AI chips and data center hardware — Polymarket · 2026-10-26
• India Needs at Least 5GW of Compute by 2028 to Stay Relevant, User Argues — EAccelerate_42 · 2026-10-26
• SpaceX AI revenue up 830%, Claude designs 318K antivenom proteins in 64 hours — PeterDiamandis · 2026-10-26
• Meta plans up to $95 billion stock sale to fund AI buildout — The Decoder · 2026-10-26
• Pandora M1: first prototype of a physical device for distributed GPU monitoring — tarantulae · 2026-10-26
• INT21's AI agent swarm rewrote Drama 3's inference runtime in Rust within one hour, beating SGLang-Omni — bingxu_ · 2026-10-26

Infra channel →