deepseek-v4.1VIDEO

DeepSeek V4.1 Flash on 4 DGX Sparks hits 107 tokens/s locally

Using MiaAI_lab's DeepSeek V4.1 Flash 1017 recipe plus a custom open-source ZCode harness, the author reports 107 tokens/s decode on four DGX Sparks and shares both repos.

Norsyx

@norsyxArchived 17 hours agoOriginally posted 2026.10.28

Norsyx @norsyxOriginally posted 2026.10.28

Open original on X

🚀 107 tokens/s on 4 DGX Sparks. Local AI goes BRR 🚀 @MiaAI_lab’s DeepSeek v4.1 Flash 1017 recipe + my open-source ZCode harness: ⚡ Decode speed + prefill stats 🌐 Network traffic logs 🐾 A mascot. Emotional support for your tokens Steal my setup Harness: https://github.com/Leoid/ZCode-patched Recipe: https://github.com/MiaAI-Lab/DeepSeek-V4.1-Flash-1017-4x-DGX-Sparks-TensorFold

x.com/norsyx/status/2115144517729050755Archived 2026.10.28 12:03