🧰 TokenSpeed

The items that name it 3

• Gunjan N Ganatra (@gnganatra) on X
  thread2026-10-14

  LLM serving engine from LightSeek Foundation, with kernel and serving optimizations; used to reach 685 output tokens/s per user on Qwen3.8-Flash-Next.

• Engrams Embedding Entendre: Codesign for Efficient DRAM/SSD Offloading
  article2026-09-18

  Serving project whose maintainers are said to collaborate with NVIDIA on day 0 CUDA optimization.

• Qwen3.8-27B (Apache-2.0, multimodal dense, 262k ctx)
  repo

  Dedicated serving engine supporting Qwen3.8 and YaRN