DeepInfer-Mac runs 750B DeepSeek on a 24 GB MacBook Air

DeepInfer-Mac runs the 750B-class DeepSeek-V4.1-Flash model on a MacBook Air with 24 GB of unified memory. The method, described in a short arxiv preprint by Haorui Li, is to park almost everything on disk and read it in on demand.

A mixed low-bit conversion of the model occupies 338 GiB on the internal SSD. About 7.6 GiB of weights stay resident. Routed experts and what the paper calls 'Engram rows' are loaded when needed. 'This separation makes the model executable, but it leaves each generated token waiting for a succession of storage reads,' Li writes.

End-to-end, the system generates English, Chinese and code at 0.21-0.72 tokens per second, with a largest sampled process footprint of 16.73 GiB. Matched decode comparisons show a 21.0% throughput gain from next-layer prefetching over resident-expert overlap alone, at the cost of additional reads. Longer prompts remain costly.

The paper describes its own operating point plainly: 'a model far larger than memory can run locally, while storage and prefill still govern the user's wait.' Two of the researchers we track posted the link in short order.