Mac mini M5 Pro 上 11 个开源模型各改写葛底斯堡演说 1000 次
在 64GB M5 Pro Mac mini 上让 11 个开源权重模型连续改写葛底斯堡演说 1000 轮：DeepSeek V4.1 Flash（343GB）与 GLM-5.3 Flash 全程保持 90% 以上相似度，靠 SSD 流式加载约 5.5 tok/s。
I made 11 open-weight models rewrite Abraham Lincoln's Gettysburg Address 1,000 times each on my Mac mini (M5 Pro, 64GB) Every rewrite fed the next one. THE WHOLE RUN LASTED FOR 45 HOURS. Here's the raw footage, timelapsed to 40 seconds. I wanted to know how many rewrites it takes before it stops being the same speech. Held on for all 1000 rounds: - DeepSeek V4.1 Flash (343GB) - GLM-5.3 Flash (88GB) Both are too big for 64GB of memory, so they streamed off the SSD at about 5.5 tok/s. After 1000 rounds they were still above 90% similar to the original. Lasted 200+ rounds: - Qwen3.8 27B: 233 - Qwen3.6 35B-A3B: 228 - Qwen3-Next 80B: 222 Faded within 100 rounds: - Qwen3.8 Flash Next (pruned): 77 - GLM-4.7 Flash: 10 Gone within 5 rounds: - Bonsai 2 9B: 5 - Gemma 4 26B: 3 - Qwen3.5 0.8B: 2 - DeepSeek-V2-Lite: 1 The smallest ones shrank the 271-word speech to 13 words. Then I turned up the randomness and Qwen3-Next 80B went from lasting 222 rounds to 34. 19,019 rewrites, all on one machine. More analysis in the thread 🧵
这条是引用着另一条写的：「Mac mini 本地连跑 11 个开源权重模型，含 DeepSeek V4.1 Flash」