2026-10-05
DeepSeek-V4.1-Flash on a 128 GB Mac
With ds4 Metal and SSD streaming, DeepSeek-V4.1-Flash in Q2 runs on an Apple Silicon Mac with 128 GB of memory, no cloud GPU needed.
DeepSeek-V4.1-Flash has a local recipe for Mac: ds4 Metal in Q2, with the weights streamed from the SSD.
Requirements
- An Apple Silicon Mac with 128 GB of unified memory.
- About 400 GB free on the SSD.
- No GPU cost: everything stays on your machine.
When to move to the cloud
For speed and long contexts there are the vLLM recipes:
- 4× B200 (TP4): about $20/h measured.
- 8× H200: about $16.5/h measured.
The cloud session starts, answers on the local endpoint and shuts down when you stop it. The Free plan gives you one hour a day to try it.