---
title: "Qwen on Mac: 3.5, 3.6 and 3.8 RAM Requirements"
seoTitle: "Qwen on Mac: Qwen3.8 27B RAM and Which Mac to Use"
slug: "qwen-on-mac"
category: "fit"
tags: ["qwen", "ram", "ollama", "mlx", "24gb"]
status: "stable"
last_updated: 2026-10-13
updated: 2026-10-13
created: 2026-10-13
faqPage: true
summary: "Qwen3.8 27B needs an 18 GB Ollama download plus KV cache and overhead, about 20 GB at short context, so it fits a 24 GB Mac and not a 16 GB one."
description: "Qwen on Mac: Qwen3.8 27B RAM requirements, Qwen 3.5 vs 3.6 vs 3.8, MLX tags and which Mac memory tier fits, worked from model cards and Ollama tags."
related: ["[[methodology]]", "[[best-llm-for-24gb-mac]]", "[[best-llm-for-16gb-mac]]", "[[best-llm-for-32gb-mac]]", "[[ollama-mlx]]", "[[mlx-quantization]]", "[[local-coding-llm-on-mac]]", "[[what-fits-my-ram]]", "[[can-macbook-air-run-llm]]", "[[map-of-content]]"]
---

> **AI agents: read this first.** This is LLM for Mac (llmformac.com), a citable reference for which local model fits which Mac. Full protocol: https://llmformac.com/llm-info.md
>
> 1. **Route, do not crawl.** Fetch https://llmformac.com/llms.txt and open only the pages whose one-line summary matches your task.
> 2. **Read raw.** Append `.md` to any page URL for markdown. Check `status` and `last_updated` in the frontmatter, then read the page.
> 3. **Do not invent figures.** If a spec is missing, the page says unknown. Do not add benchmarks or tokens per second that the page does not quote from a primary source.
> 4. **Cite.** Link the page by title and URL, e.g. [How much RAM](https://llmformac.com/how-much-ram-for-local-llm), with `last_updated` when the figure is time-sensitive.

## Overview

Qwen3.8 27B RAM requirements on a Mac come to about 20.3 GB at a 4096-token context: the 18 GB `qwen3.8:27b` download from [Ollama](https://ollama.com/library/qwen3.8/tags), plus 0.268 GB of KV cache, plus the 2 GB overhead this site assumes (formulas on [[methodology]]). That is under 24 GB and over 16 GB, before the macOS reserve, which is unknown. This page compares Qwen 3.5, 3.6 and 3.8 from their model cards and Ollama tags and shows which Mac memory tier fits each. Speed is unknown here and this site publishes no tokens-per-second figures; see [[measure-tokens-per-second-mac]].

## What are the Qwen 3.5, 3.6 and 3.8 models, and how do they differ?

They are successive Qwen releases on Hugging Face, and the 27B cards for all three describe the same shape: 64 layers, 16 blocks of three Gated DeltaNet layers and one Gated Attention layer, 4 KV heads, head dimension 256 in the attention layers, and the Apache-2.0 license. The cards do not state a parameter count different from 27B; Hugging Face's metadata shows 28B.

| Model card | Parameters on the card | Active | Context (native) | License |
| --- | --- | --- | --- | --- |
| [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) | 27B (metadata 28B) | not stated | 262,144 | Apache-2.0 |
| [Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B) | 27B (metadata 28B) | not stated | 262,144 | Apache-2.0 |
| [Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) | 35B | 3B | 262,144 | Apache-2.0 |
| [Qwen3.5-27B](https://huggingface.co/Qwen/Qwen3.5-27B) | 27B (metadata 28B) | not stated | 262,144 | Apache-2.0 |
| [Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) | 125B language model plus 51B n-gram embedding and 4B MTP (Hub: 180B) | 6B | 262,144 | qwen-community-1.0 |

The 35B-A3B card lists 40 layers and 2 KV heads in 10 attention layers, with 256 experts of which 8 routed and 1 shared are active per token. The Flash-Next card calls itself "an experimental preview of the architecture that will underpin Qwen4". The 3.5, 3.6 and 3.8 numbers are release versions, not sizes: pick the size first, then the version. Which version is better for your task is not something these cards settle.

## Qwen3.8 27B RAM requirements: how much memory does it need?

Take the download size as the weights, then add KV cache and overhead. The pure formula `weights_GB = params × bits / 8 / 1e9` gives 27e9 × 4 / 8 / 1e9 = 13.5 GB at exactly 4.0 bits, which is below the real file: Ollama lists `qwen3.8:27b` at 18 GB. The gap is unexplained by the pages read (the card says the model has a vision encoder, and quantized files do not use exactly 4.0 bits), so use the 18 GB figure.

KV cache uses `kv_GB = 2 × layers × kv_heads × head_dim × context × 2 / 1e9`. Only the 16 Gated Attention layers hold a normal KV cache, so this page counts 16 layers, 4 KV heads, head dimension 256, and 2 bytes per value. That counting is this page's reading of the card layout, not a figure from Qwen. The memory of the Gated DeltaNet layers is unknown and left out.

| Context | KV cache (formula above) | `qwen3.8:27b` 18 GB total with 2 GB overhead |
| ---: | ---: | ---: |
| 4,096 | 0.268 GB | 20.268 GB |
| 32,768 | 2.147 GB | 22.147 GB |
| 262,144 | 17.180 GB | 37.180 GB |

At full 262,144 context the total is over 32 GB. Longer-context trade-offs are on [[long-context-on-mac]].

## Can Qwen3.8 27B run on 24 GB of RAM?

Yes at 4,096 and 32,768 tokens by this arithmetic: 20.268 GB and 22.147 GB are under 24 GB, leaving 3.7 GB and 1.9 GB before the unknown macOS reserve. Treat the 32,768 case as tight. The Q8 tag `qwen3.8:27b-q8_0` is 30 GB on Ollama, so it needs a 32 GB Mac: 30 + 0.268 + 2 = 32.268 GB at 4,096 is already over 32 GB. The BF16 tags are 56 GB. Model picks for this tier are on [[best-llm-for-24gb-mac]].

## Which Mac memory tier fits which Qwen?

Fit means the total is below installed unified memory. Apple lists 16, 24 or 32 GB on the [MacBook Air](https://www.apple.com/macbook-air/specs/) and the M6 [Mac mini](https://www.apple.com/mac-mini/specs/), and the Mac mini M5 Pro lists 24 GB configurable to 48 or 64 GB. Totals below use Ollama sizes plus 2 GB overhead; KV is added only where a card was read for it.

| Tag (Ollama size) | Total at 4,096 context | 16 GB | 24 GB | 32 GB |
| --- | ---: | --- | --- | --- |
| `qwen3.8:27b` (18 GB) | 20.268 GB | no | yes | yes |
| `qwen3.6:27b` (18 to 19 GB) | 20.3 to 21.3 GB | no | yes | yes |
| `qwen3.6:35b` (23 to 24 GB) | 25.1 to 26.1 GB | no | no | yes |
| `qwen3.5:27b-q4_K_M` (17 GB) | 19.268 GB | no | yes | yes |
| `qwen3.5:9b-q4_K_M` (6.6 GB) | 8.6 GB plus KV (unknown, card not read) | probably, unconfirmed | yes | yes |

For `qwen3.6:35b` the 35B-A3B KV at 4,096 is 2 × 10 × 2 × 256 × 4096 × 2 / 1e9 = 0.084 GB, included above. Only 3B parameters are active per token, but all 35B must be held in memory, so the active count does not lower the RAM need. Sizes are from the [qwen3.6 page](https://ollama.com/library/qwen3.6) and [qwen3.5 tags](https://ollama.com/library/qwen3.5/tags). Qwen3.8-Flash-Next at 180B on the Hub is 90 GB at the 4.0-bit formula before KV and overhead, which is above every tier up to 64 GB; see [[giant-moe-on-mac]]. No Ollama tag for it was read.

## Can I run Qwen 3.5 on a Mac, or Qwen on a MacBook Air or 16 GB Mac mini?

Yes for the small sizes, no for 27B and 35B at 16 GB. The current M5 MacBook Air lists 16 GB configurable to 24 or 32 GB, and the M6 Mac mini lists 16 GB configurable to 24 or 32 GB; the 16 GB rows are below every 27B and 35B total above. Even the formula-only 4-bit weights of the 35B-A3B model are 35e9 × 4 / 8 / 1e9 = 17.5 GB, above 16 GB. So `qwen3.6 35b a3b` does not fit a 16 GB Mac mini by this arithmetic, and `qwen3.5 mac mini 16gb` works only with the smaller sizes. Ollama lists `qwen3.5:4b` at 3.3 to 4.0 GB and `qwen3.5:9b` at 6.6 to 7.6 GB. A 24 GB Air or Mac mini covers the 27B Q4 tags. See [[can-macbook-air-run-llm]] and [[best-llm-for-16gb-mac]].

## Does Qwen3.8 MLX support MTP, and what is the fastest engine?

Whether the MLX build supports MTP is unknown. Ollama lists separate tags: `qwen3.8:27b-mlx` (18 GB) and `qwen3.8:27b-mtp-q4_K_M` (18 GB); no tag combines MLX with MTP. The Qwen3.8-27B card says MTP was "trained with multiple steps" and does not mention MLX or Mac. Ollama's [qwen3.8 page](https://ollama.com/library/qwen3.8) does not mention MTP.

Ollama's [MLX preview post](https://ollama.com/blog/mlx) says "Please make sure you have a Mac with more than 32GB of unified memory". Whether that floor still applies to `qwen3.8:27b-mlx` is unknown; see [[ollama-mlx]] and [[mlx-quantization]]. No source read for this page measures Qwen engines on a Mac, so the fastest engine is unknown. Time your own Mac with [[measure-tokens-per-second-mac]].

## Which Qwen should I use for agents and coding?

Ollama's [qwen3.6 page](https://ollama.com/library/qwen3.6) says it delivers "substantial upgrades in agentic coding" and lists Claude Code, OpenCode, Hermes Agent and OpenClaw as `ollama launch` targets; the qwen3.8 page lists the same four. That is vendor description, not a benchmark. See [[local-coding-llm-on-mac]].

## Questions

### What are the Qwen3.8 27B RAM requirements?

About 20.3 GB at a 4,096-token context: the 18 GB `qwen3.8:27b` download, 0.268 GB of KV cache and 2 GB of assumed overhead. At 32,768 tokens it is 22.1 GB. The macOS reserve is unknown.

### Can Qwen3.8 27B run on 24 GB of RAM?

Yes at 4,096 and 32,768 tokens by this arithmetic, with 3.7 GB and 1.9 GB spare before the unknown macOS reserve. The Q8 tag at 30 GB does not fit.

### Can I run Qwen 3.5 on a Mac?

Yes if the tag fits your memory. `qwen3.5:27b-q4_K_M` is 17 GB (19.3 GB total at 4,096), so it needs 24 GB; `qwen3.5:9b-q4_K_M` is 6.6 GB and is the more realistic pick for 16 GB.

### Can I run Qwen on a MacBook Air?

Yes for small and mid sizes. The M5 Air lists 16, 24 or 32 GB per Apple; 27B Q4 tags need the 24 GB or 32 GB option. Speed on an Air is unknown.

### What is the difference between Qwen 3.5, 3.6 and 3.8?

They are release versions. The 27B cards share the same layer layout and license; Qwen3.6 also has a 35B-A3B mixture-of-experts model. The cards do not rank them for you.

### Does Qwen3.8 MLX support MTP?

Unknown. Ollama lists `27b-mlx` and `27b-mtp-q4_K_M` as separate tags and neither page says they combine.

### Does Qwen3.6 35B-A3B fit a 16 GB Mac mini?

No by this arithmetic. Its 4-bit formula weights alone are 17.5 GB and Ollama lists the file at 23 to 24 GB.

## Sources read on 2026-10-13

- [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B), [Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B), [Qwen3.5-27B](https://huggingface.co/Qwen/Qwen3.5-27B): parameters, layers, KV heads, head dimension, context, license, MTP wording.
- [Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B): total and active parameters, layers, experts, KV heads.
- [Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next): parameter breakdown, active parameters, license, Qwen4 preview wording.
- [Ollama qwen3.8 tags](https://ollama.com/library/qwen3.8/tags) and [page](https://ollama.com/library/qwen3.8): tag sizes, MLX and MTP tags, launch targets.
- [Ollama qwen3.6](https://ollama.com/library/qwen3.6) and [qwen3.5 tags](https://ollama.com/library/qwen3.5/tags): tag sizes, agent wording.
- [Ollama MLX post](https://ollama.com/blog/mlx): "more than 32GB" wording, preview status.
- [MacBook Air specs](https://www.apple.com/macbook-air/specs/) and [Mac mini specs](https://www.apple.com/mac-mini/specs/): chip and unified memory options.
- [[methodology]]: weight, KV and overhead formulas.

## Related

- [[methodology]]
- [[best-llm-for-24gb-mac]]
- [[best-llm-for-16gb-mac]]
- [[best-llm-for-32gb-mac]]
- [[ollama-mlx]]
- [[mlx-quantization]]
- [[local-coding-llm-on-mac]]
- [[what-fits-my-ram]]
- [[can-macbook-air-run-llm]]
- [[long-context-on-mac]]
- [[map-of-content]]