DeepSeek-V4.1-Flash Benchmarked on NVIDIA DGX Spark Boxes

Benchmarks of DeepSeek-V4.1-Flash across four desktop NVIDIA DGX Spark boxes reveal key architectural details of the model, which holds approximately 748 billion parameters in total. Despite its massive size, the model utilizes an asymmetric encoder-decoder that activates only 8 billion parameters during prefill and 16 billion during decode. Testing on vLLM demonstrated strong performance across prefill, single-user, and concurrent user benchmarks. Open article →