DeepSeek V4.1 Flash 在 4 台 DGX Spark 上 91.9 分钟造出音乐机
用 EXL3 4.0bpw 量化、TP4 拆分到四台 DGX Spark，DeepSeek Harness 一次提示完成音乐机，浏览器评分 10/10，代码解码 82.8 tok/s。
Community Day, post 6: the third world. DeepSeek-V4.1-Flash-1017 built a music machine on four DGX Sparks in 91.9 min, run on @bertholomusai's TensorFold recipe. bertholomus' own EXL3 4.0 bpw quant, one model split across all four Sparks (TP4). Agent: DeepSeek Harness 0.2.1-rc.2. One prompt, one try, no help. Thinking on, T 1.0, top-p 0.95, the recipe's own sampling. Music machine: done in 91.9 min Tokens written: 327,264 Browser check: 10.0/10 Decode, code: 82.8 tok/s (separate speed run, thinking off) Decode, prose: 80.8 tok/s Context: 1,048,576 tokens, the recipe's setting The model named it AETHERIUM, a brass-and-glass marble orchestra. Marbles hit bells, harp strings, marimba bars, pipes and drums, and every hit plays a note. Type "happy 8-bit chase" and it plays that piece. Open: it ran on the previous version of my music-machine prompt (a text field plus presets). I changed the prompt today. 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: 34,446 tokens. My Sparks are cabled as a switchless ring, so it runs in the recipe's neighbor-ring mode, not the switched setup the author serves. At 90 min the test tells the agent to finish. This one stopped at 91.9 min. The music machine was bertholomus' idea, thank you. Play the music machine and the voxel world in your browser. Every number, the exact command and the raw output: https://niklaslenz.com/v2/deepseek-v41-flash-4sparks-bertholomus-tp4/ Recipe, quant and TensorFold fork by @bertholomusai. TensorFold engine by @ashxhart. Model by DeepSeek. My pick on 4 Sparks: DeepSeek-V4.1-Flash-1017 on bertholomus' recipe.