On this page
Qwen 3.8 27B runs on any Apple silicon Mac with 16 GB of memory or more. Off Grid AI Desktop runs it for free and offline: Bonsai 2 27B (7.2 GB plus a 0.63 GB vision file) on a 16 GB Mac, a 4-bit file of about 15 GB on 24 GB, and the full 4-bit download (17.4 GB) on 32 GB.
An 8 GB Mac can’t hold any build of the 27B. Run Qwen 3.5 9B there instead.
Download Off Grid AI Desktop (free, no account)
What fits your Mac
Off Grid AI Desktop checks every download against your Mac’s memory. A model up to 65% of your memory Runs well. Up to 82% it’s Tight, which means it loads but leaves little room for long chats. Past that it’s Too big.
Here’s what that rule gives for Qwen 3.8 27B. Sizes include the 0.93 GB vision file (0.63 GB for Bonsai 2).
| Your Mac | File to use | Size | What the app shows |
|---|---|---|---|
| 8 GB | None. Use Qwen 3.5 9B | n/a | Too big |
| 16 GB | Bonsai 2 27B | 7.8 GB | Runs well |
| 24 GB | 4-bit UD-IQ4_XS | 15.3 GB | Runs well |
| 24 GB | 4-bit UD-Q4_K_M (the app’s default) | 17.4 GB | Tight |
| 32 GB | 4-bit UD-Q4_K_M (the app’s default) | 17.4 GB | Runs well |
| 32 GB | 5-bit UD-Q5_K_M | 20.8 GB | Runs well |
| 36 GB | 6-bit UD-Q6_K | 22.9 GB | Runs well |
| 48 GB or more | 8-bit Q8_0 | 30.0 GB | Runs well |
File sizes come from Unsloth’s Qwen 3.8 27B files and Prism ML’s Bonsai 2 27B. The verdicts are the app’s own rule applied to those sizes.
What Qwen 3.8 27B is
Qwen 3.8 27B is Alibaba’s smallest open Qwen 3.8, released in August 2026 under the Apache 2.0 licence. It’s a dense model that reads text, images and video. It thinks before it answers by default, and it handles 262,144 tokens of context natively.
On a Mac, the app runs it through llama.cpp on Metal, so your Mac’s GPU does the work. Apple silicon shares one pool of memory between the CPU and GPU, which is why the memory size is the number that matters.
Why 16 GB uses Bonsai 2
The standard 4-bit Qwen 3.8 27B is 16.5 GB on its own, so it can’t fit in 16 GB with macOS running. Bonsai 2 27B is Prism ML’s build of the same model with every weight stored as -1, 0 or +1. That takes it from about 54 GB to 7.2 GB.
Prism ML says it keeps 98.2% of the full model’s score across 14 benchmarks, against 72.59 points for a standard 2-bit file of a similar size. Off Grid AI Desktop ships Prism’s engine for it, so there’s nothing to build.
How to set it up
- Download Off Grid AI Desktop, install it and open it.
- Open Models > Text.
- Pick the file for your Mac:
- 16 GB: download Bonsai 2 27B.
- 32 GB or more: download Qwen 3.8 27B. That’s the 4-bit file.
- 24 GB: click the Qwen 3.8 27B card to open its panel. Under Sizes, pick a 4-bit file that doesn’t say Tight, such as UD-IQ4_XS. All files shows every size.
- Wait for the model and its vision file to finish. It’s 7.8 to 17.4 GB, so use a fast connection.
- Open a new chat, select the model and ask something you can check.
Try a task that shows what a 27B is good at:
Here’s a function that should remove duplicate emails from a list, but it keeps the wrong one when two differ only in capital letters. Explain the bug and fix it.
Then paste the code. The first reply takes longer than a small model’s, because Qwen 3.8 thinks first.
To ask about a picture, attach an image to the chat. Then turn off Wi-Fi and ask again. It keeps answering, because nothing leaves your Mac.
Keep it fast
- Leave memory free. Close heavy apps while it answers. A browser with many tabs competes for the same memory.
- Start a new chat per topic. A long history uses memory on every reply.
- Lower the context if it struggles. Open the model’s settings and reduce Context window. Most chats don’t need the full 262,144 tokens.
- Don’t force a Tight file. It loads, but long chats are where it runs out of room.
If something goes wrong
| Symptom | Check | Fix |
|---|---|---|
| The card says Too big | Your Mac’s memory | Use the file from the table above, or Qwen 3.5 9B |
| The card says Tight | File size against memory | Pick a smaller size in Sizes, or use it for short chats only |
| Replies get slow in long chats | Context window | Lower it, or start a new chat |
| Your Mac slows down while it answers | Other apps using memory | Close them, or go one size down |
| Image questions fail | The vision file | Check it finished downloading |
| Download stops | Free disk space | Free up at least 20 GB for the 4-bit file |
Need the full-size model?
On a 16 GB Mac, Bonsai 2 is how the 27B fits. If you’d rather not squeeze it onto your Mac at all, Off Grid AI Boost is a flat-price API for open models on our GPUs, and Qwen 3.8 is on it. It’s from $4.99 a month for the first cohort, and we don’t store your prompts or train on them. Boost starts in January; pre-order to lock in half price. Usage limits aren’t published yet.
FAQ
Can a 16 GB Mac run Qwen 3.8 27B? Yes, as Bonsai 2 27B. It’s Prism ML’s 7.2 GB build of Qwen 3.8 27B, plus a 0.63 GB vision file, and Off Grid AI Desktop lists it for 16 GB. The standard 4-bit file is 16.5 GB and is too big for a 16 GB Mac.
What’s the best Qwen 3.8 27B file for a 24 GB Mac? A smaller 4-bit file of about 14 to 15 GB, such as UD-IQ4_XS. The app’s default 4-bit download is 17.4 GB with its vision file, which the app marks Tight on 24 GB.
Does Qwen 3.8 27B run well on a 32 GB Mac? Yes. The 4-bit download Off Grid AI Desktop lists (UD-Q4_K_M, 17.4 GB with the vision file) is well inside what the app counts as Runs well on 32 GB.
Do I need an internet connection? Only for the download. After that, Qwen 3.8 27B runs on your Mac with Wi-Fi off, and your chats stay on your disk.
Is there a smaller Qwen 3.8, like a 9B? Not from Alibaba. The 4B and 9B Qwen 3.8 files on Hugging Face are community distills. Off Grid AI Desktop points you to Qwen 3.5 4B and 9B instead.
Related guides
- How to Run Qwen 3.8 Privately on a 16GB Laptop in 2026
- What Mac Do You Need to Run Qwen 3.8 or DeepSeek in 2026?
- How to Run Qwen 3.6 35B-A3B on Your Mac in 2026
- How to Run Qwen 3.5 on Your Mac in 2026
Download Off Grid AI Desktop and run Qwen 3.8 27B on your Mac today.