Pranav Innovates

You commented GIANT. Here’s everything from the reel.

How big an AI can a 16 GB Mac run?

From @PranavInnovates (the "Comment GIANT" reel, Nov 2026). I ran the 8B and 35B rungs myself on a MacBook Air (M1, 16 GB, macOS 15). The 26B and 125B numbers are other people's published runs, and they're labelled as theirs.

The trick behind the big ones: these are mixture-of-experts models. Only a few of the experts are used for each token, so these tools keep a small core in memory and read just the experts they need from the SSD.

8B: Ollama + Llama 3.1, the usual way

ollama.com · ollama run llama3.1:8b

  • The 4.9 GB model is loaded whole into memory. ollama ps showed 6.2 GB in use on my Air.
  • Fast, easy, and the right choice for models that fit.

26B: TurboFieldfare (Google's Gemma 4 26B-A4B)

github.com/drumih/turbo-fieldfare · Apache-2.0 · 7.3K ★

  • Streams experts from the SSD and runs in about 2 GB of RAM.
  • Its community benchmarks list an M1 MacBook Air with 16 GB at 6.1 to 7.8 tokens a second, depending on prompt size.
  • Needs macOS 26 (Metal 4), so I couldn't run it on mine.

35B: Swiftlet (Qwen3.6-35B-A3B), the one I ran

github.com/leonickson1/Swiftlet · Apache-2.0

Terminal
git clone https://github.com/leonickson1/Swiftlet && cd Swiftlet
swift build -c release          # needs Swift 6.2 (Xcode 26, or a swift.org toolchain)
.build/release/swiftlet-repack --from-hf Leonickson/Qwen3.6-35B-A3B-qpack --output ~/models/qwen3.6-35b.qpack   # 18 GB
.build/release/swiftlet chat ~/models/qwen3.6-35b.qpack --cache-gb 2 "Write a haiku about a tiny laptop running a giant AI."
  • My run: the model is 18 GB on disk, peak memory was 3.2 GB (macOS footprint, GPU buffers included), the first word came after 6.2 s, and it wrote at 2.1 tokens a second (2.6 to 3.2 in other runs).
  • The haiku: "Small screen, vast mind hums, / Silicon dreams fill the space, / World in palm, awake."
  • It's slow: fine for short answers. One user's 591-token prompt took 265 s to the first token on an M1. A newer chip is much faster (7 to 11 tokens a second on an M5, per the README).
  • You need about 20 GB of free disk.

125B: Slotstream (Qwen3.8-Flash-Next)

github.com/carloslfu/slotstream · MIT

  • An independent run (github.com/atbender/qwen38-flash-next-mac-mini) on a 16 GB M1 Mac mini with macOS 26: 1.76 tokens a second, 5.09 GB peak memory.
  • The model is about 105 GB on disk and needs about 110 GB free, so I didn't try it.

More free guides: pranavinnovates.com