deepseek-v4.1POST

DeepSeek V4.1 Flash on two DGX Sparks: 2,008 tok/s prefill at 128K, 113 tok/s decode

Local deployment with BLACKFOLD on two DGX Sparks reports 2,008 tok/s prefill at 128K context and 113 tok/s code decode, with a rerun planned.

Wayne Lowry

@wikiwayneArchived 10 days agoOriginally posted 2026.10.21

Wayne Lowry @wikiwayneOriginally posted 2026.10.21

Open original on X

Shoutout to my boy @blackwellboy 🔥 DeepSeek-V4.1-Flash 1017 running locally on two DGX Sparks with BLACKFOLD: 2,008 tok/s prefill at 128K context and 113 tok/s code decode. That's the kind of receipt local AI needs. Going to try a rerun on my GX10… when my second Spark arrives!

This post quotes another post: x.com/blackwellboy/status/2112728704023757080

x.com/wikiwayne/status/2112880328004288756Archived 2026.10.22 12:01