Serving DeepSeek-V4-Flash on 4x DGX Spark: the 3.76.1 recipe

An in-house vLLM image serving DeepSeek-V4-Flash at ~71 tok/s decode, 400K context, thinking enabled by default — distilled from a month of benchmarking, three confirmed bugs, and one configuration audit. Extracted verbatim from production.

• DeepSeek-V4-Flash
• vLLM
• multi-node
• TP4
• DGX Spark
• FP8
• MoE
• thinking mode
• long context

• time~45 min per node, plus 15–20 min first boot
• yields4-node cluster serving DSV4-Flash at ~71 tok/s, 400K context, thinking on by default