DeepSeek V4.1 Flash-1017 on 8x DGX Spark cluster: 111 tok/s at C4

Running the official FP8 weights on an 8-node DGX Spark cluster with vLLM and TP8 gives 81 tok/s at C1, 103 at C2 and 111 at C4 on the default recipe.

FPGA des GARÇONS

@hypermac6502Archived 11 days agoOriginally posted 2026.10.18

DeepSeek V4.1 Flash-1017 on 8x DGX Spark cluster: 111 tok/s at C4 (image 1 of 4)

DeepSeek V4.1 Flash-1017 on 8x DGX Spark cluster: 111 tok/s at C4 (image 2 of 4)

DeepSeek V4.1 Flash-1017 on 8x DGX Spark cluster: 111 tok/s at C4 (image 3 of 4)

DeepSeek V4.1 Flash-1017 on 8x DGX Spark cluster: 111 tok/s at C4 (image 4 of 4)

FPGA des GARÇONS @hypermac6502Originally posted 2026.10.18

Open original on X

DGX Spark x8 Cluster : deepseek-ai/DeepSeek-V4.1-Flash-1017 C1 : 81 tok/s C2 : 103 tok/s C4 : 111 tok/s vLLM: Eugr’s spark-vllm-docker DGX Spark: TP8 tool-eval-bench: v2.7.1 This is the official frontier FP8 model. Default recipe,It just works🔥

x.com/hypermac6502/status/2111623407351775730Archived 2026.10.18 16:26

Open original on X