两台 DGX Spark 跑 DeepSeek V4.1 Flash：70.6 tok/s，附完整跑分

Niklas @niklaslenz_ai原推发布于 2026.10.25

在 X 上打开原推

🚨New contender for the two Spark category🚨 DeepSeek-V4.1-Flash-1017 on two DGX Sparks: 70.6 tok/s with @bertholomusai's TensorFold recipe. bertholomus' own EXL3 2.5 bpw quant, one model split across both Sparks (TP2). The recipe v0.7.2 as of Oct 24, commit 30a5bf9, with the author's sampling and context. Thinking off for every number here, except the GPQA line marked with thinking. The recipe wants RDMA GID index 5, which my Sparks did not have, so I gave the link ports IPv4 addresses. One Spark, an ASUS GX10, has an older driver. Decode, code: 70.5 tok/s Decode, prose: 70.6 tok/s 500 words: about 10.0 s Reading a 32K-token prompt: 1,830 tok/s First token after a 32K prompt: 17.9 s Four requests at once: 107.4 tok/s in total Eight requests at once: 106.3 tok/s in total Context: 1,048,576 tokens, the recipe's setting Largest context I tested: 131,072 tokens The recipe runs 4 slots, so with eight requests the rest wait for a free slot. For scale, bertholomus' 4.0 bpw quant on four Sparks: Decode, code: 82.8 tok/s Decode, prose: 80.8 tok/s Eight requests at once: 141.2 tok/s in total MMLU-Pro: 62.0% ± 4.7 points GPQA Diamond: 43.9% ± 6.9 points IFEval: 92.0% ± 3.8 points LiveCodeBench: 60.0% ± 14.5 points Long text 8K: 67.5% ± 14.0 points Long text 32K: 57.5% ± 14.7 points GPQA with thinking: 80.0% ± 13.9 points Open: MMLU-Pro and GPQA answers were capped at 48 tokens (one letter). Answers that hit a cap count as wrong: MMLU-Pro: 49 of 400 cut GPQA Diamond: 46 of 198 cut Long text: 3 answers cut LiveCodeBench: 2 answers cut GPQA with thinking: 5 of 30 cut Quality and tool calls in Mia's sparkDash 2.50.0: QA: 89.3% ± 5.0 points MMLU: 93.0% ± 3.0 points Tool Eval: 93/100 Every number, the exact command and the raw output: https://niklaslenz.com/v2/deepseek-v41-flash-2sparks-bertholomus-tp2/ You can play the world it built there in your browser. Recipe, quant and TensorFold fork by @bertholomusai. TensorFold engine by @ashxhart. sparkDash by @MiaAI_lab. Model by DeepSeek.