DeepSeek V4.1 Flash on two DGX Sparks: 63.6 tok/s and full benchmarks
Deploys DeepSeek V4.1 Flash across two DGX Sparks with sudoingX's vLLM recipe and EXL3 2.8bpw quant (TP2), hitting 62–65 tok/s decode and 1,797 tok/s on a 32K prompt, with MMLU-Pro 71.3 and LiveCodeBench 65.0.
A new contender for the 256 class with DeepSeek-V4.1-Flash-1017 on two DGX Sparks: 63.6 tok/s with @sudoingX's vLLM recipe. sudoingX's own EXL3 2.8 bpw quant, one model split across both Sparks (TP2) on vLLM, with the DSpark drafter from his pack. His release v2.0 as of Oct 22, commit f8f20b3. Sampling from the DeepSeek model card (temperature 1.0, top_p 1.0), thinking off for every number here, except the GPQA line marked with thinking. My cable plugs into different network ports on the two Sparks, so my launcher copy takes one network interface per node. The recipe's settings are unchanged. Decode, code: 64.9 tok/s Decode, prose: 62.2 tok/s 500 words: about 11.6 s Reading a 32K-token prompt: 1,797 tok/s First token after a 32K prompt: 18.2 s Four requests at once: 66.2 tok/s in total Eight requests at once: 66.3 tok/s in total Context: 262,144 tokens, the recipe's setting Largest context I tested: 131,072 tokens MMLU-Pro: 71.3% ± 4.4 points GPQA Diamond: 52.5% ± 6.9 points IFEval: 90.0% ± 4.2 points LiveCodeBench: 65.0% ± 14.2 points Long text 8K: 70.0% ± 13.7 points Long text 32K: 55.0% ± 14.7 points GPQA with thinking: 80.0% ± 13.9 points The recipe runs 2 slots, so with four or eight requests the rest wait for a free slot. Open: MMLU-Pro and GPQA answers were capped at 1,024 tokens. 4 of 400 and 16 of 198 hit the cap and count as wrong. 3 long-text answers and 5 of 30 GPQA answers with thinking also ran past our token limits. Play the voxel world it built in your browser. Every number, the exact command and the raw output: https://niklaslenz.com/v2/deepseek-v41-flash-2sparks-sudoingx-vllm/ Recipe, vLLM patches and EXL3 quant by @sudoingX. DeepSeek-V4.1-Flash-1017 by @deepseek_ai.