四台 DGX Spark 本地跑 V4.1 Flash 对比云端 Opus 5.5 建体素世界
同题各跑一次：Opus 5.5 云端 33.2 分钟、约 7.82 美元，四台 Spark 上的 V4.1 Flash 49.7 分钟、零 API 费用，浏览器检查同为 9.0/10。
Claude Opus 5.5 in the cloud against DeepSeek-V4.1-Flash on four DGX Sparks: same voxel-world prompt, one try each. Opus built Lumenfall in my Claude Code CLI. The Sparks run @bertholomusai's TensorFold recipe (his EXL3 4.0 bpw quant split across all four Sparks, commit 8f84e7a) and built Aetherfall. Opus 5.5, cloud: done in 33.2 min DeepSeek, 4 Sparks: done in 49.7 min Tokens written, Opus: 173,159 Tokens written, DeepSeek: 121,707 Browser check, Opus: 9.0/10 Browser check, DeepSeek: 9.0/10 Cost, Opus: about $7.82 at API list prices (Anthropic pricing page, Oct 23, 2026), I ran it on my subscription Cost, 4 Sparks: no API bill, my own hardware The worlds are really not that far apart when it comes to quality. The setups differ: the Sparks ran in the locked test container, answer limit 262,144 tokens, in the recipe's neighbor-ring mode on my switchless ring. Opus ran in my own CLI with internet, my files, no time gate and about 71,000 tokens of context before the first answer. Times run from agent start (Sparks) and from sending the prompt (Opus; 26.0 min by the model's own clock file, which started late). Play the Spark world in your browser. Every number, the exact command and the raw output: https://niklaslenz.com/v2/deepseek-v41-flash-4sparks-bertholomus-tp4/ New generations from Qwen, DeepSeek and GLM are due in the next days or weeks. I'm curious what their worlds will look like. Recipe, quant and TensorFold fork by @bertholomusai. TensorFold engine by @ashxhart. Model by DeepSeek. Claude by Anthropic. My pick: Lumenfall from the cloud is a better world, but if it has to run at home, bertholomus' 4-Spark ring gives you Aetherfall at the same browser-check score with no API bill.