deepseek-v4.1VIDEO

DeepSeek V4.1 Flash 1017 NVFP4 on 4 DGX Sparks: 92–98 tok/s single stream

NVFP4 deployment across four DGX Sparks in a QSFP ring beats vLLM on 1–3 streams, hitting 92–98 tok/s counting and 85–89 tok/s code, with a 114K-token task at 62.5 tok/s.

Carlos Alvarado

@calvarado2004Archived 9 days agoOriginally posted 2026.10.23

DeepSeek V4.1 Flash 1017 NVFP4 on 4 DGX Sparks: 92–98 tok/s single stream (image 1 of 2)

DeepSeek V4.1 Flash 1017 NVFP4 on 4 DGX Sparks: 92–98 tok/s single stream (image 2 of 2)

Carlos Alvarado @calvarado2004Originally posted 2026.10.23

Open original on X

DeepSeek-V4.1-Flash-1017 NVFP4 on NVIDIA Flare, across 4 DGX Sparks in a QSFP ring, no switch. Faster than vLLM on 1-3 Streams. 1 stream: 92–98 tok/s count, 85–89 code. Faster than our vLLM recipe. One-shot neon city at Max effort: 114K tokens at 62.5 tok/s. Recipe: https://github.com/calvarado2004/deepseek-v41-1017-nvfp4-nvidia-flare @MiaAI_lab @Tech2Wild @mmastrac @Libertai_DAI

x.com/calvarado2004/status/2113342060887433562Archived 2026.10.23 12:08