llama.cpp 合并 V4.1 Flash 1017 修复：M2 Ultra 上 30.8 tok/s

Steven | SelfHostedSI @SelfHostedSI原推发布于 2026.10.27

在 X 上打开原推

DeepSeek retired V4 Flash Vision Exp on its API: calls to that name now get V4.1 Flash 1017. On Sunday Georgi Gerganov merged a llama.cpp fix he made with that same model, running on a 192GB M2 Ultra. His PR logs the run: 28.1 minutes, prompt at 215 tok/s, output at 30.8 tok/s, 39 lines added across 4 files. The weights are still on Hugging Face, smallest GGUF 98.6 GB. On your own disk, a model only changes when you swap the file.