Deepseek 4 flash running locally on a high-end Mac
could run directly on a Mac with an M3 Max chip and 96 GB of memory. The setup used Antirez’s special engine, a ds4 gguf file, and the `--ssd-streaming` option. It also used `iogpu.wired_limit_mb=86016` to give Metal more available memory.
Changing the code to raise the cache safety value to 0.70 may allow more expert parts of the model to be loaded into . The measured generation speed was about 11 to 13 . From a in an empty Jan assistant chat, startup took about 10 seconds; after that, was about 3 to 5 seconds.
Longer input processing was still frustratingly slow, so day-to-day usefulness was uncertain. A simple menu bar daemon was also built in about 20 minutes so the server could be launched from Spotlight.
Key points
- ran on an M3 Max Mac with 96 GB of memory.
- The setup used Antirez’s engine, a ds4 gguf file, and `--ssd-streaming`.
- Generation speed was roughly 11 to 13 .
- took about 10 seconds, then was about 3 to 5 seconds.
- Longer input processing remained slow, which may limit agent use.