Which Qwen to Run

Every model below is open-weight and verified on Alibaba’s Hugging Face organization. Speeds are the LLMCheck index’s figure for each Mac, refilled on every build; the mark beside each one shows whether it is published, community-measured or estimated.

ModelScoreOllama tag (download)MacM5 MaxMac mini M6
Qwen 3.6-35B-A3B
MoE, 3B active
81qwen3.6:35b-a3b (~24 GB)24 GB tight, 32 GB+community130 tok/scommunity59 tok/s
Qwen3.8-27B
dense, image + video
72qwen3.8:27b (18 GB)24 GB+community31 tok/scommunity10 tok/s
Qwen 3.6-27B
dense
70qwen3.6:27b (18–19 GB)24 GB+estimated~30 tok/sestimated~9 tok/s
Qwen3-Coder-Next
MoE coder, 3B active
70qwen3-coder-next (52 GB)64 GBestimated~84 tok/sdoes not fit
Qwen3.8-Flash-Next
MoE, 6B active
67qwen3.8-flash-next:125b-mlx (105 GB)128 GBcommunity47 tok/sdoes not fit
Qwen 3.5 9B
dense
61qwen3.5:9b (6.6–7.6 GB)16 GBestimated~81 tok/sestimated~30 tok/s

The short version: run Qwen 3.6-35B-A3B unless you have a reason not to. It ranks highest because only 3B of its 35B parameters work on each token, so it is three to four times faster than the dense 27B models at similar quality. Pick Qwen3.8-27B when answer quality matters more than speed: it is the strongest dense model in the index by capability (46/50) and reads images and video. Ollama tags and sizes were checked on ollama.com on 15 November 2026.

Step 1: Check Your Memory & Pick Your Model

The single most important factor is unified memory. To check yours: click the Apple menu → About This Mac, and read the Memory line. macOS keeps part of that memory for itself, so leave a few gigabytes beyond the model’s download size.

Shopping for a Mac to run Qwen? The sweet spot is 32 GB of unified memory: it runs Qwen 3.6-35B-A3B with room to spare. The best model by RAM page shows what each memory tier unlocks, and the Mac hardware hub compares configurations.

Step 2: Install Ollama

Ollama is the easiest way to run Qwen on a Mac. Download it from ollama.com, open the .dmg, and drag Ollama into your Applications folder. Launch it once so it installs its command-line tool, then open Terminal and check the version:

ollama --version

You want 0.40 or later (0.40.16, released 9 November, was the current stable release on 14 November 2026). From 0.40, Ollama runs model architectures that MLX supports on MLX by default on Apple Silicon (release notes), so you get Apple’s own framework without extra setup. For the full walkthrough with troubleshooting, follow the Install Ollama on Mac guide first, then come back here.

Step 3: Pull & Run Qwen

A single command downloads your chosen model and drops you into a chat:

# Best pick for 32 GB Macs (24 GB: Recursive Flint) — MoE, 3B active (about 24 GB download)
ollama run qwen3.6:35b-a3b

# Strongest dense model, reads images and video (18 GB download)
ollama run qwen3.8:27b

# Dense Qwen 3.6 (18–19 GB download)
ollama run qwen3.6:27b

The first run pulls the model, so expect a wait on broadband; after that it launches in seconds. When you see the >>> prompt, you are talking to Qwen entirely on your own machine — nothing leaves your Mac. Type /bye or press Ctrl+D to leave. To download without starting a chat, use ollama pull with the same tag.

According to the LLMCheck index, here is what to expect for generation speed (each figure carries its provenance mark and is refilled on every build):

ModelM5 MaxM5 ProMac mini M6
Qwen 3.6-35B-A3Bcommunity130 tok/scommunity79 tok/scommunity59 tok/s
Qwen3.8-27Bcommunity31 tok/ssourced17 tok/scommunity10 tok/s
Qwen 3.6-27Bestimated~30 tok/sestimated~15 tok/sestimated~9 tok/s

The dense models feel sluggish on base and Pro chips; the 35B-A3B stays fast everywhere it fits. The Best by Mac pages list the figure for your exact configuration.

Step 4: Prefer an App? Run It in LM Studio

If you would rather skip the Terminal, LM Studio (free) runs the same models behind a graphical interface:

  1. Download LM Studio from lmstudio.ai and drag it into Applications.
  2. Open the Discover tab and search for Qwen3.6 or Qwen3.8.
  3. Pick an MLX 4-bit or GGUF Q4_K_M build of the model you chose in Step 1 (LM Studio flags builds that will not fit).
  4. Click Load, then chat in the Chat tab.

LM Studio also shows live tok/s and memory pressure, which makes it a good way to see what your Mac can handle before committing to a workflow. Full setup details are in the LM Studio on Mac guide.

Step 5: Use It in Your Apps via the API

Ollama serves an OpenAI-compatible API at http://localhost:11434/v1, so any tool that speaks the OpenAI format can use Qwen as a local backend. A quick test with curl:

curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.6:35b-a3b",
    "messages": [{"role": "user", "content": "Write a haiku about Apple Silicon."}]
  }'

The same endpoint works from Python with the official OpenAI SDK — point base_url at localhost:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:11434/v1",
    api_key="ollama"  # required by the SDK, ignored locally
)

resp = client.chat.completions.create(
    model="qwen3.6:35b-a3b",
    messages=[{"role": "user", "content": "Summarize the plot of Dune in 3 sentences."}]
)
print(resp.choices[0].message.content)

That makes Qwen a free, private stand-in for cloud APIs in scripts, agents, RAG pipelines and editor extensions. LM Studio exposes an equivalent server on port 1234.

Step 6: Tune It — Quantization, Context & MLX

Quantization

The default tags are 4-bit, the best balance of quality, speed and memory for most Macs. With memory to spare, an 8-bit build is a small quality step up:

# 8-bit: more memory, a little slower
ollama pull qwen3.6:27b-q8_0      # 30 GB
ollama pull qwen3.6:35b-a3b-q8_0

Eight-bit makes sense on 48 GB and up. The quantization guide covers the trade-offs.

Context length

Longer contexts use more memory. To raise the window for a session, set num_ctx:

>>> /set parameter num_ctx 32768

On a 24 GB Mac, keep context modest (8K–16K) to avoid swapping. With 48 GB or more you can push much further for long-document work.

MLX directly

Ollama 0.40 already uses MLX for these models. To drive Apple’s framework yourself — for scripting, or to try other builds — install mlx-lm:

pip install -U mlx-lm

mlx_lm.generate \
  --model mlx-community/Qwen3.6-35B-A3B-4bit \
  --prompt "Explain mixture-of-experts in two sentences."

The mlx-community 4-bit builds are 20 GB for the 35B-A3B and 16 GB for each 27B. MLX matters most if you would otherwise run GGUF files through llama.cpp: in a published comparison on an M4 Max, an MLX-native server ran Qwen 3 30B-A3B at about 110 tok/s against about 90 for llama.cpp (arXiv 2601.19139). The MLX guide goes deeper.

128 GB Macs: Qwen3.8-Flash-Next

Qwen3.8-Flash-Next is Alibaba’s preview of the Qwen 4 architecture: 125B parameters with 6B active, plus a 51B n-gram embedding that has to sit in memory, which is why its 4-bit builds are 102–112 GB. According to the LLMCheck index it runs at community47 tok/s on a 128 GB M5 Max. In Ollama, run ollama run qwen3.8-flash-next:125b-mlx (105 GB). Two other routes work:

# llama.cpp (support merged 27 August 2026) — 102 GB IQ4_NL GGUF
llama-server -hf ggml-org/Qwen3.8-Flash-Next-GGUF:IQ4_NL

# MLX — needs mlx-vlm 0.7 or later (stock mlx-lm support is still a pull request)
pip install -U mlx-vlm
mlx_vlm.generate --model mlx-community/Qwen3.8-Flash-Next-4bit \
  --prompt "Summarise this repository" --max-tokens 512

Read its licence first: Qwen Community 1.0 allows local and internal use but requires a separate commercial licence for model-as-a-service and coding or office assistant products. Details in Qwen3.8-Flash-Next on a Mac.

That’s it. You now have the LLMCheck index’s best Qwen for your Mac running locally, for free and fully private. When Qwen 4 weights appear, the release tracker and this guide will say how to run them.