How much RAM does Qwen3.8 27B need?

About 20.4 GB at a 32,768-token context: the 17.6 GB file (Unsloth's UD-Q4_K_XL quantization) plus 2.8 GB of buffers, measured from llama.cpp's own log on a 32 GB Mac mini M4. PocketWebTools for Mac files it under 32 GB Macs.

Does Qwen3.8 27B run on a 16 GB Mac?

No. The file alone is 17.6 GB. On a 16 GB Mac the app offers Qwen3.5 9B, a 6.0 GB download that replied at 19 tokens per second on our 16 GB M1 Pro.

How fast is Qwen3.8 27B on a Mac?

5.3 to 5.8 tokens per second when replying and about 56 tokens per second when reading a prompt, on a 32 GB Mac mini M4 with every layer on the GPU. It loads in about 20 seconds and a follow-up in the same chat starts in about 0.6 seconds. Another Mac can be faster or slower: speed depends mostly on the chip, and the app shows your own speed under every reply.

Qwen3.8 27B or Qwen3.6 35B-A3B?

Both are filed under 32 GB Macs. On the same 32 GB Mac mini M4, Qwen3.6 35B-A3B replied at 25 to 27 tokens per second against 5.3 to 5.8 and scored 9 of 9 on our test against 8.5, so the app suggests it first. Qwen3.8 27B is the smaller file, 17.6 GB against 22.4 GB, and needs about 3 GB less memory.

Does Qwen3.8 27B think before it answers in the app?

No. Qwen's model card says thinking is on by default; the app starts each reply after an empty think block, so the answer begins straight away.

Can I use Qwen3.8 27B commercially?

Yes. It is released under Apache 2.0, which allows commercial use. The app ships the license text and adds no restrictions of its own.

Qwen3.8 27B: RAM, Speed, Settings

17.6GB

Download

20.4GB

Memory once loaded

5 to 6tokens/s

Reply speed, Mac mini M4

32Ktokens

Context window in the app

Which Mac runs it

The chat models the app ships, by the memory your Mac has. Loaded is the file plus the buffers llama.cpp allocates at the app's context window. Speeds are from a 16 GB M1 Pro; the two 32 GB models ran on a 32 GB Mac mini M4. Speed depends mostly on the chip, not the memory size, so another Mac can be faster or slower than ours. The app shows the speed of every reply on your own Mac.

Model	Mac memory	Download	Loaded	Reply speed
Qwen3.8 27BThis page	32 GB and up	17.6 GB	20.4 GB	5.3 to 5.8 tokens/sMac mini M4
Qwen3.6 35B‑A3B	32 GB and up	22.4 GB	23.7 GB	25.1 to 27.3 tokens/sMac mini M4
Qwen3.5 9B	16 GB and up	6.0 GB	7.1 GB	18.6 to 19 tokens/sM1 Pro
Gemma 4 12B	16 GB and up	7.4 GB	8.8 GB	13.8 to 14.7 tokens/sM1 Pro
Bonsai 27B (1‑bit)	8 GB and up	3.8 GB	5.1 GB	14 to 16.3 tokens/sM1 Pro
Qwen3.5 2B	8 GB and up	1.3 GB	2.0 GB	73 to 76 tokens/sM1 Pro

See which one we would pick for your Mac

Qwen3.8 27B speed on a 32 GB Mac mini M4

Measured in the app's own engine: llama.cpp on Metal, every layer on the GPU, other apps closed. Speed depends mostly on the chip, not the memory size, so another Mac can be faster or slower than ours. The app shows the speed of every reply on your own Mac.

Load, file already on disk: 20.3 s
Reply speed: 5.3 to 5.8 tokens/s
Qwen3.6 35B‑A3B replied at 25 to 27 tokens per second on the same Mac.
Reading a prompt: About 56 tokens/s
A 774-token prompt took 13.8 s before the first word appeared.
Follow-up in the same chat: 0.59 s to the first word
452 of 483 prompt tokens came from the cache.
Our nine-prompt test: 8.5 of 9
Summarizing, rewriting, questions about an 800-token memo, list formatting, TypeScript and a reasoning puzzle. The half point lost: its summary called a planned bus link a rail link.

On a MacBook Pro M5 Pro, measured by Artificial Analysis

Not our run. Artificial Analysis timed a recorded agent task (168 turns, 172 tool calls) on their own MacBook Pro with an M5 Pro, 64 GB and a 20-core GPU, with thinking on and MTP speculative decoding, neither of which the app uses. Compare the two models with each other here, not with our speeds above.

Qwen3.8 27B: 37.6 min to finish
Q4_K_M file, llama.cpp b11026 on Metal.
Qwen3.6 35B‑A3B: 11.9 min to finish
UD-Q4_K_M file, llama.cpp b11026 on Metal.
Intelligence Index: 34 for Qwen3.8 27B, 18 for Qwen3.6 35B‑A3B
Artificial Analysis's score for the model itself, the same on any Mac. Higher is better.

Source: Artificial Analysis, laptop and workstation inference, read October 24, 2026.

Settings the app runs Qwen3.8 27B with

The same values work in any llama.cpp front end.

File: Qwen3.8-27B-UD-Q4_K_XL.gguf
Unsloth's 4‑bit dynamic quantization, pinned to one revision and checked against its SHA-256 after download.
Context window: 32,768 tokens
GPU layers: All, on Metal
Batch size: 512
Temperature: 0.7
Top-p / top-k / min-p: 0.9 / 40 / 0.05
Thinking: Off
The reply starts after an empty think block, so the answer begins straight away.
Prompt cache: Kept between turns
A follow-up only reads the new message, not the whole chat again.

What the app does with Qwen3.8 27B

Chat

Threads saved on your Mac, replies rendered as Markdown with code blocks and tables, and a meter that shows how much of the 32K window a thread has used.

Clip

Reads a video's transcript and picks the moments worth cutting. The app picks with it only when neither 16 GB model nor Qwen3.6 35B‑A3B is installed, since those reply faster.

Extract, then ask

Send to Chat hands the text of a scan or PDF to the chat screen. The 32K window holds a long document in one go.

Where it falls short

Every speed here comes from one 32 GB Mac mini M4. The 8 and 16 GB models on these pages were timed on a 16 GB M1 Pro, an older chip, so the two sets are not directly comparable.
It is the slowest chat model in the lineup: about a fifth of the reply speed of Qwen3.6 35B‑A3B on the same Mac, and a long pasted prompt takes several seconds to read.
It is filed under 32 GB Macs. On our 32 GB Mac, macOS let one app use 26.8 GB of graphics memory and this model takes 20.4 GB. The app checks before loading and refuses a model that does not fit, so with other apps open it may ask you to close them first.
The model can read images, but the app runs it on text only: it does not load the image part of the model.
Nine prompts is a small test. It checks everyday writing and reading jobs; it is not a leaderboard.

Run Qwen3.8 27B in PocketWebTools for Mac

One app for chat, transcription, documents, voices, photos and video, with every model on your Mac. $99 once, free lifetime updates.

Buy for $99$49

Already have a key? Download the app. It asks for the key the first time it opens.

14-day money-back guarantee. Your license key is shown right after checkout and stays on your purchase page; one key works on up to 3 Macs.

See everything in the Mac app

Qwen3.8 27B questions

How much RAM does Qwen3.8 27B need?: About 20.4 GB at a 32,768-token context: the 17.6 GB file (Unsloth's UD-Q4_K_XL quantization) plus 2.8 GB of buffers, measured from llama.cpp's own log on a 32 GB Mac mini M4. PocketWebTools for Mac files it under 32 GB Macs.
Does Qwen3.8 27B run on a 16 GB Mac?: No. The file alone is 17.6 GB. On a 16 GB Mac the app offers Qwen3.5 9B, a 6.0 GB download that replied at 19 tokens per second on our 16 GB M1 Pro.
How fast is Qwen3.8 27B on a Mac?: 5.3 to 5.8 tokens per second when replying and about 56 tokens per second when reading a prompt, on a 32 GB Mac mini M4 with every layer on the GPU. It loads in about 20 seconds and a follow-up in the same chat starts in about 0.6 seconds. Another Mac can be faster or slower: speed depends mostly on the chip, and the app shows your own speed under every reply.
Qwen3.8 27B or Qwen3.6 35B‑A3B?: Both are filed under 32 GB Macs. On the same 32 GB Mac mini M4, Qwen3.6 35B‑A3B replied at 25 to 27 tokens per second against 5.3 to 5.8 and scored 9 of 9 on our test against 8.5, so the app suggests it first. Qwen3.8 27B is the smaller file, 17.6 GB against 22.4 GB, and needs about 3 GB less memory.
Does Qwen3.8 27B think before it answers in the app?: No. Qwen's model card says thinking is on by default; the app starts each reply after an empty think block, so the answer begins straight away.
Can I use Qwen3.8 27B commercially?: Yes. It is released under Apache 2.0, which allows commercial use. The app ships the license text and adds no restrictions of its own.