DeepSeek Flash moves to new cluster: throughput jumps from 189 to 241 tok/s
teortaxesTex · x · 2026-10-29
A long-time DeepSeek API watcher noticed the model fingerprint of the new Flash checkpoint changed, indicating Flash was moved to another cluster with noticeably faster inference.
Session stats shared over a month apart show TPS rising from 189 to 241 tok/s, while average TTFT slightly worsened from 2.9s to 3.3s. The author speculates DeepSeek's load has grown dramatically and it may be "batch-maxxing," and jokes that recent slowness suggested it wasn't running on Blackwell GPUs.
Related event: DeepSeek Flash Appears to Move to New GPU Cluster with Faster Inference(2 posts)→
More from Infra
- GPU availability on neoclouds vanishes as queue times grow, founder warns ladder is being pulled up — generativist · 2026-10-29
- Analyst: interconnect supplier market ripe for consolidation around SerDes IP — BenBajarin · 2026-10-29
- Apple can outgrow PC market by 8-10% as edge token generators, says analyst — BenBajarin · 2026-10-29
- Alpine Linux discloses GitLab security incident over attempted CVE-2026-97229 exploitation — sashk · 2026-10-29
- exe.dev Ships WIF Support: Agents Make OAuth Token Exchange Finally Usable for All — davidcrawshaw · 2026-10-29
- AI pushes on GitHub up 4.9x in a year; the bottleneck isn't the platform, it's human code review — FinanceYF5 · 2026-10-29