Mac Studio M5 Ultra 本地跑 V4.1 Flash：65 分钟建体素世界
LSQ3 量化 + oMLX，一次提示生成体素世界，浏览器检查 8/10，写入 16.8 万 token，解码 63.8 tok/s，128K 上下文下 45.1 tok/s。
256 GB Mac class: DeepSeek-V4.1-Flash-1017 on a Mac Studio M5 Ultra built a voxel world in 65.7 minutes. Recipe and LSQ3 quant by gnahzhang, oMLX engine by @jundotkim, model by @deepseek_ai. Build test: the model works as an agent. One prompt, one try, no help. Browser check: 8/10 Tokens written: 168,228 Speed benchmark: 63.8 tok/s decode (separate run, thinking off) Agent settings: model-card sampling, temperature 1.0 and top-p 1.0. Answer limit: raised to the model card's 262,144 tokens. The recipe's server default is 16,384. My test gate sent a keepalive every 30 s so the connection stayed open. The browser check ran on a second machine, the Mac's Docker VM has too few CPU cores for it. Same page files. 137 of 400 MMLU-Pro answers ran past our short answer limit and count as wrong. 149 of 198 GPQA answers did the same. Memory: tested up to 128K of the 1M window, decode there 45.1 tok/s. Close to the 2-Spark world, on one Mac. Play the world in your browser. Every number, the exact command and the raw output: https://niklaslenz.com/v2/deepseek-v41-flash-mac-lsq3-g128-omlx/