Outmake Beta Download Beta
Guides

AI video on a 16 GB Mac

Google's AI summary for running Wan 2.2 on a Mac recommended 36 to 64 GB of memory or more when we read it on 13 October 2026. The 14B model doesn't need that, if the engine reads the model from disk a part at a time instead of loading it whole. These are the measurements.

The model

Wan 2.2 14B at its smallest size (Q4_K_S) is a 21 GB download as Image to Video and 23 GB as Text to Video: two halves of nearly 9 GB each, one for the early steps and one for the late ones, plus a 3.7 GB text reader. Loaded whole, the two halves alone are more than a 16 GB Mac has.

What was measured

On 5 October 2026, on a Mac with an M1 Pro and 16 GB, the same short video (320 by 320, 17 frames) was made four times, each with a different limit on how much of the model the engine may hold at once.

LimitTimeHeldLeast memory free
9.6 GB186 s8.8 GB10%
7 GB198 s7.1 GB15%
5 GB208 s5.1 GB22%
3 GB211 s3.5 GB32%

Giving the model a third of the memory cost an eighth more time. At 9.6 GB the Mac had a tenth of its memory free. So Outmake sets the limit to 6.5 GB on a 16 GB Mac: about twice the room left for everything else, for about 6% longer by its own estimate from these rows. The engine holds a little over the limit, as the Held column shows, and the video itself takes about another gigabyte.

Bigger videos take longer

The table is a small video. On an M1 Pro with 14 graphics cores, Wan 2.2 14B took 150 to 172 seconds a step at 464 by 464 over 33 frames (4 October), and 200 seconds a step at 624 by 352 over 33 frames (5 October). Wan 2.2 5B, which fits in memory whole, took 37 seconds a step at 864 by 480 over 49 frames.

It is a video made on a laptop with nothing sent anywhere.

Clips made this way, with the minutes each took

What hasn’t been measured

  • No Mac with less than 16 GB has run it. Outmake’s limits there (3 GB under 12 GB of memory, 4.5 GB under 16) are what the table says would fit.
  • From 24 GB up the limit is 0.6 of memory, where a whole half stays loaded and nothing is read twice.
  • These are from one Mac.

In Outmake

Open Models and choose the Video tab. Each video model says how it fits this Mac and how long a video takes before you download it. Download one, open Pictures, choose Video, and type what you want. The limit is set for your Mac’s memory; there is nothing to configure.

In stable-diffusion.cpp yourself

Outmake makes video with stable-diffusion.cpp. Reading the model in parts is two of its flags:

--params-backend disk --max-vram 6.5

The number is in GiB. Without the second flag, the engine keeps every part it has read, and on 16 GB both halves end up in memory and it runs out.

Wan 2.2 14B Text to Video on every Mac · Image to Video · Wan 2.2 5B