Building
I ran a 27B model on a 16GB MacBook Air. Here's what it's actually good for.
A post went viral on X this week about PrismML's Ternary Bonsai 2 27B: a Qwen-based model squeezed down to 1.75 bits per weight, so the whole thing is a 5.9 GB file. PrismML says it keeps about 98% of the full model's quality. An independent tester on Hugging Face got closer to 87% on harder tasks.
So I tested it myself on the laptop I actually own: an M2 MacBook Air with 16GB of RAM, running PrismML's own build of llama.cpp. (The download alone took about three and a half hours.)
Speed: slow, but it works
- About 8 tokens per second while generating, dropping to 6 or 7 on long answers.
- Time to first answer: 2 seconds for a tiny prompt, about 55 seconds for a 2,200-token prompt.
- A logic puzzle took 8 to 13 minutes per try.
Memory: you'll feel it
The model peaked at 7.3 GB, but the Mac's swap went from 2.8 GB to 12.75 GB. You can run it, but you won't enjoy using the laptop for anything else while it runs.
Accuracy: better than I expected
I ran the same tasks with thinking off and with medium thinking:
| Task | Thinking off | Medium thinking |
|---|---|---|
| Messy email to 17 fields | 16/17 | 17/17 |
| Call transcript to 11 fields | 11/11 | 11/11 |
| Tagging 12 messages | 12/12 | 12/12 |
| Tool calls | 4/5 | 5/5 |
| Coding (unit tests passed) | 21/23 | 22/23 |
| Logic puzzle | 5/5 | 5/5 |
On easy, practical jobs it landed a lot closer to PrismML's claim than to the skeptic's.
The weak spot: dates
It booked "tomorrow at 3pm" for today. It gave the wrong weekday in a reply. It wrote 2025 when the prompt clearly said 2026. If a model like this is scheduling anything, check the date in code. Don't trust the model on it.
Verdict
It's not a daily driver: too slow and too memory-hungry for anything a person sits and waits on. But as an overnight worker on a spare Mac, sorting messages and pulling facts out of emails and transcripts, it's genuinely useful, and it costs nothing per token.
The interesting part isn't that a 27B model fits on a 16GB laptop. It's that the boring jobs most businesses actually need (pull these fields, tag these messages, call this tool) already work at this size.
Links: the viral post, model card, independent test.