双 DGX Spark 本地跑 V4.1 Flash：88 分钟生成体素世界
用 Mia 的 EXL3 2.9bpw 量化在两张 DGX Spark 上跑 V4.1 Flash，DeepSeek Harness 一次提示生成体素世界耗时 88.1 分钟、写 15.9 万 token，代码解码 72 tok/s，浏览器评分 8.5/10。
Community Day, post 8: the fourth world. DeepSeek-V4.1-Flash-1017 on two DGX Sparks built this voxel world in 88.1 min, on the new release of @MiaAI_lab's 2-Spark recipe. Her EXL3 2.9 bpw quant on TensorFold, one model split across both Sparks. Her release from Oct 21, commit 370daf0. Agent: DeepSeek Harness 0.2.1-rc.2, DeepSeek's own agent. One prompt, one try, no help. Thinking on. Sampling from Mia's recipe: temperature 1.0, top_p 0.95. Voxel world: done in 88.1 min Tokens written: 159,465 Browser check: 8.5/10 Decode, code: 72.0 tok/s (separate speed run, thinking off) Decode, prose: 65.0 tok/s Context: 524,288 tokens, the setting Mia measures with AURELIA, The Last Lantern: floating islands at dusk, a lighthouse, a sky whale and its own camera tour, all in one HTML file. Open: DeepSeek Harness shortened two old tool outputs once and wrote one summary of the earlier work between 82.7 and 84.1 min of the 88.1. The agent kept building afterwards. Answer limit in the build test: 262,144 tokens, the figure on the DeepSeek model card. Mia's server default is 32,768, so I raised it. The recipe's default context is 1,048,576 tokens. Longest answer: 28,397 tokens, so none was cut. During the build test, a log command of mine used one CPU core on the head Spark for 2 minutes. Play the world in your browser. Every number, the exact command and the raw output: https://niklaslenz.com/v2/deepseek-v41-flash-2sparks-mia-tensorfold/ Recipe and EXL3 quant by @MiaAI_lab. TensorFold engine by Ash Hart, its DeepSeek-V4.1 support by BertholomusAI, model by DeepSeek. My pick on 2 Sparks: DeepSeek-V4.1-Flash-1017 on @MiaAI_lab's recipe.