双 DGX Spark 跑 V4.1 Flash 体素世界：73.7 分钟、解码 70 tok/s
用 EXL3 2.5bpw 量化在两张 DGX Spark 上 TP2 部署，一次提示生成体素世界耗时 73.7 分钟、写出 13.8 万 token，浏览器检查 8.0/10，代码解码 70.5 tok/s。
DeepSeek-V4.1-Flash-1017 built this voxel world on two DGX Sparks in 73.7 min, run on @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. Agent: DeepSeek Harness 0.2.1-rc.2. One prompt, one try, no help. Thinking on, T 1.0, top-p 0.95, top-k 20, the recipe's own sampling. Voxel world: done in 73.7 min Tokens written: 138,675 Browser check: 8.0/10 Decode, code: 70.5 tok/s (separate speed run, thinking off) Decode, prose: 70.6 tok/s 500 words: about 10.0 s Eight requests at once: 106.3 tok/s in total Context: 1,048,576 tokens, the recipe's setting LUMENFALL: floating isles around a world tree, with lanterns, airships, an observatory, a sunken temple, a crystal spire, waterfalls and its own camera tour. I like the art style of DeepSeek a lot and you can recognize it here. For scale, the same prompt on bertholomus' 4.0 bpw quant on four Sparks: Voxel world: done in 49.7 min Browser check: 9.0/10, scored with the older capture method Open: the browser check took 2 points off, for the orbit-by-mouse test and the frame rate in the test's software rendering. Frame rate there: 0.2 fps Scene size: 2 million triangles The check's closing screenshot stalled, so I re-ran the same checks with a different capture method. Answer limit in the build test: 262,144 tokens, the figure on the model card. The recipe's default is 32,768, so I raised it. Longest answer: 46,330 tokens, so none was cut. Prompt at the end: 190,984 tokens. 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. At 90 min the test tells the agent to finish. This run was done before that. Play the world in your browser. Every number, the exact command and the raw output: https://niklaslenz.com/v2/deepseek-v41-flash-2sparks-bertholomus-tp2/ Recipe, quant and TensorFold fork by @bertholomusai. TensorFold engine by @ashxhart. Model by DeepSeek.