DeepSeek-1017 Grows Past Two Sparks: Official Weights on Four

Four Sparks fit the official weights. Three already post numbers.

@MiaAI_lab showed DeepSeek V4.1 Flash 1017 running on four DGX Sparks: official weights on TensorFold, with eight concurrent streams and thinking on. “Important fact: official weights,” she adds. @ashxhart, who builds TensorFold, posted the same run.

She has not published speeds. She asked a builder running her recipe with native weights on SGLang across four Sparks to run sparkDash decode benchmarks: thinking on, temperature 0.7, one to eight streams.

On three Sparks, @laky_fx posted sparkDash results for DeepSeek 4.1 Flash on TensorFold 1.2.5 with tensor parallel 3. One stream ran 126.1 tok/s on code and 87.7 tok/s on prose. Twelve code streams ran 246.2 tok/s in total. Four Sparks have an advantage, he says: an extra box and a cleaner split of the model.