Qwen3.6 on Mac: 27B, 35B-A3B, Vision and Ollama

Newer version: For the updated release, see our Qwen3.8-27B analysis.

Qwen3.6 is useful for Mac users only if you keep the variants separate. There is a dense Qwen3.6-27B checkpoint, the Qwen3.6-35B-A3B MoE model and several Ollama tags with different package sizes. Qwen’s model cards and the current Ollama page list the family with image input.¹²³

Four questions matter: which Qwen3.6 variant is a sensible Apple-Silicon starting point, which Ollama commands are currently listed, how the tags differ, and which benchmark scores actually belong to which model.

---

Qwen3.6: model facts and Ollama tags (October 2026)

The model cards specify 262,144 native context tokens and describe YaRN as an optional route to longer contexts; Ollama shows the package size and modality for each currently listed tag.¹²³

Criteria	Qwen3.6-27B	Qwen3.6-35B-A3B
Release	April 22, 2026	April 14, 2026
Architecture	Dense	Mixture of Experts
Parameters	27B	35B total / 3B active
Active per token	all 27B	3B active
Vision	yes, with vision-capable tag/runtime	yes
Context	262,144 native, up to ~1,010,000 extended	262,144 native, up to ~1,010,000 extended
License	Apache 2.0 / Open Weights	Apache 2.0 / Open Weights
Strength	Dense profile, vision, local Ollama use	MoE, agentic coding, vision
Mac suitability	18–19 GB Ollama package; little reserve on 24 GB	23–24 GB Ollama package; more headroom is sensible

---

Qwen3.6-27B vs Qwen3.6-35B-A3B — Not the Same

Qwen3.6-27B is dense: most parameters take part in each generated token. Its 17–19GB quantized Ollama packages may be a more practical first experiment than 35B-A3B, but no independent Apple Silicon stability or speed measurement is provided here.

Qwen3.6-35B-A3B is a Mixture-of-Experts model: 35B total parameters, but only 3B activate per token. It is more efficient at inference than a dense model of the same total size, but it requires more care around tag selection, runtime, context length and vision support.

---

Which Ollama Tag Should I Use on Mac?

Per the Ollama library and tag list (rechecked October 9, 2026):³⁴

Goal	Recommended tag	Why
Vision + local use	qwen3.6:27b	18–19 GB, Text + Image, lower barrier to entry
Larger vision/MoE variant	qwen3.6:35b	23–24 GB, Text + Image; 32 GB experimental, 48 GB+ leaves usable headroom
MLX variant	qwen3.6:27b-mlx	19 GB, MLX, Text + Image
MoE/MLX variant	qwen3.6:35b-mlx	24 GB, MLX, Text + Image
Maximum quality / BF16	BF16 tags	only for very large Macs/workstations

Ollama currently lists all four compared tags as Text, Image. That is a statement from the model page, not a guarantee of identical image quality or memory use in every runtime. Check the concrete tag and runtime before relying on it.³⁴

ollama run qwen3.6 uses the current default/latest tag. If you specifically want the lighter local starting point, name the tag explicitly: qwen3.6:27b.

---

Setup with Ollama

Install Ollama on macOS

Use the official macOS installer: download the Ollama DMG, move the app to Applications and open it once. The running app exposes the local server and offers to make the ollama CLI available in your shell. Installing a CLI alone is not proof that a server is running. Ollama currently requires macOS 14 Sonoma or later; Apple Silicon supports CPU/GPU execution.

ollama --version
curl --fail-with-body -sS http://127.0.0.1:11434/api/tags

A connection error means the local service is not responding. Start the Ollama app before attempting model downloads; do not expose 11434 on 0.0.0.0 as a shortcut. Ollama macOS, networking FAQ.

Download a model

# Vision-capable 27B entry point (Text + Image)
ollama pull qwen3.6:27b

# Larger 35B-A3B/MoE variant (Text + Image)
ollama pull qwen3.6:35b

# MLX tag — Ollama lists Text + Image
ollama pull qwen3.6:27b-mlx

# MoE/MLX tag — Ollama lists Text + Image
ollama pull qwen3.6:35b-mlx

Start a model

# Vision variant
ollama run qwen3.6:27b

# MLX variant
ollama run qwen3.6:27b-mlx

Vision: Analyze Local Files with Ollama

Ollama lists qwen3.6:27b, qwen3.6:35b and the two -mlx tags with text and image inputs. That support label does not imply accurate OCR.³⁴

Private screenshots do not need to be uploaded to public websites. The official Ollama vision documentation says the Python SDK accepts local file paths, URLs or bytes. The raw REST API expects Base64-encoded image data. After pip install ollama:

from pathlib import Path
from ollama import Client

client = Client(host="http://127.0.0.1:11434")
image_path = Path.home() / "Desktop" / "test-image.jpg"
if not image_path.is_file():
    raise FileNotFoundError(image_path)

response = client.chat(
    model="qwen3.6:27b-q4_K_M",
    messages=[{
        "role": "user",
        "content": "Describe this screenshot. Flag uncertain text.",
        "images": [str(image_path)],
    }],
)
print(response.message.content)

The script reads an image from your Mac and sends it to the explicitly selected loopback Ollama server. A public image URL or external AI account is unnecessary. To verify that inference itself is local, download a non-cloud tag, turn off Ollama Cloud features and repeat a prompt with the network disconnected. External agent plugins are separate data paths. Ollama FAQ.

Floating Default Tag vs Pinned Q4 Quantization

Ollama’s tag listing shows 18–19 GB for qwen3.6:27b and 17 GB for qwen3.6:27b-q4_K_M. These are different tags, and a download size is not total runtime RAM. For repeatable checks:

ollama pull qwen3.6:27b-q4_K_M
ollama show qwen3.6:27b-q4_K_M
ollama run qwen3.6:27b-q4_K_M

Record the actual model digest for strict reproducibility. 256K context is a model capability, not a sensible default for a 24 GB or 32 GB Mac. Start at 4K or 8K and inspect memory pressure.

---

Thinking Used Correctly — Without the Thinking-Trace Hype

Qwen3.6 is trained for thinking workflows. But you should not blindly force long thinking traces:

• For normal responses, a brief explanation is enough.
• For agents, preserve_thinking is more important — the model uses previous thinking/work contexts more consistently.
• enable_thinking: false disables thinking on supported runtimes.

In Ollama without special API parameters: simply ask for a brief explanation and the result, rather than asking the model to output its full reasoning chain.

---

Benchmarks: Which Numbers Belong to Which Model?

Many values use special agent scaffolds, long context windows or tool setups — not directly transferable to your Mac.

Qwen3.6-35B-A3B, per Qwen (April 2026)

The following values come from Qwen’s own publication and are vendor benchmarks, not Mac measurements.⁵

Benchmark	Score	Interpretation
AIME 2026	92.7 %	full AIME 2026 I & II, per Qwen
MMLU-Pro	85.2 %	knowledge/reasoning benchmark
LiveCodeBench v6	80.4 %	coding
SWE-bench Verified	73.4 %	agentic coding, internal scaffold
Terminal-Bench 2.0	51.5 %	5 runs, 256K context, special harness
MMMU	81.7 %	vision/multimodal
MathVista mini	86.4 %	visual mathematical reasoning

Qwen3.6-27B

The 27B values come from the separate Qwen3.6-27B publication, so they are not mixed with the 35B-A3B table.⁶ Qwen uses special scaffolds, long context windows, multiple runs and tool setups for agent evaluations. The scores help with orientation, but they are not Mac measurements.

---

RAM / Unified Memory Recommendations

The Ollama model size is only a rough indicator. Actual memory use depends on quantization, context length, KV cache, runtime and other apps.

Unified memory	Recommendation
16 GB	Smaller Qwen3 models or heavily quantized 27B experiments; Qwen3.6 is not a comfortable fit
24 GB	17–20 GB tags leave little reserve; short context and low background load
32 GB	Practical 27B test; 35B-A3B remains experimental and context-limited
48 GB+	Much more room for 35B-A3B, vision and longer context
64 GB+	More headroom for large tags and context, still runtime-dependent

Speed on Mac depends heavily on Mac model, RAM, context, quantization, Ollama/MLX version, prompt length and vision.

---

Qwen3.6 vs Alternatives on Mac

Model	Local on Mac?	Vision?	Strength	Note
Qwen3.6-27B	yes, quantized realistically	yes with right tag	Coding, vision, dense 27B all-rounder	needs 24-32 GB to be comfortable
Qwen3.6-35B-A3B	yes, but prefer 32 GB+	yes	MoE, agentic coding, long contexts	check tags/quantization carefully
Gemma 3 27B	yes	yes	Vision + solid local quality	older, different benchmarks
Qwen3 14B/32B	yes	mostly text-only depending on tag	lighter/faster	less agent focus than Qwen3.6
Cloud models	no/local not	varies by provider	maximum quality	privacy/cost/API

---

Reproducible local API check: context, timing and memory

Before choosing between 27B and 35B-A3B, verify the exact quantization, the local backend, and the context your Mac can actually allocate. Ollama’s FAQ warns that concurrent requests multiply the context allocation; an agent with several workers can run out of headroom even if one chat worked. The following makes one local, non-streaming call with 8K context to a pinned 27B Q4 tag and prints provider-reported generation metrics. It runs locally only after you execute it; all benchmarks shown derive strictly from verified documentation rather than unrecorded local runs.

ollama pull qwen3.6:27b-q4_K_M
ollama list
ollama show qwen3.6:27b-q4_K_M

curl --fail-with-body -sS http://127.0.0.1:11434/api/chat \
  -H 'Content-Type: application/json' \
  -d '{"model":"qwen3.6:27b-q4_K_M","messages":[{"role":"user","content":"Explain why all MoE weights still occupy memory."}],"stream":false,"options":{"num_ctx":8192,"temperature":0}}' \
  | python3 -c 'import sys,json; r=json.load(sys.stdin); print("model:",r.get("model"),"output tokens:",r.get("eval_count"),"output tok/s:",round(r["eval_count"]*1e9/r["eval_duration"],2) if r.get("eval_duration") and r.get("eval_count") else "n/a","cold load s:",round(r.get("load_duration",0)/1e9,2))'

ollama ps

Interpretation: eval_count / (eval_duration / 1e9) measures generated-token throughput only, not first-token latency, image decoding, end-to-end response time or task accuracy. Run the same prompt multiple times after warm-up; report both cold load and warm-run median separately. Record the model digest, quantization, num_ctx, macOS/Ollama release, processor allocation (ollama ps) and memory pressure. Do not call vendor SWE-bench results a Mac benchmark.

Local-only is more than a localhost URL

Ollama binds to 127.0.0.1:11434 by default. If you handle private screenshots or repository code, the documented disable_ollama_cloud: true setting in ~/.ollama/server.json disables its hosted models and web search after restart. The local API address alone is insufficient evidence if a cloud-tagged model was chosen. Inspect browser/MCP/agent tools separately, and never expose an unauthenticated Ollama port over a LAN. Ollama FAQ.

What to verify before loading a 20GB+ model

The package size is not the full runtime footprint. On a 24GB Mac, a 27B tag around 18–19GB leaves limited headroom once macOS, KV cache and other applications are counted. The 35B-A3B tag is a different model; fewer active experts do not mean that most of its 35B weights can be omitted from memory.

Run ollama show qwen3.6:27b to inspect the tag you actually downloaded and ollama ps while it is running. Start with a short context and the real workflow you care about. Test image handling explicitly if vision is required; the registry label is not a measured vision-quality guarantee.

Scope: This is a Qwen3.6 setup guide. Qwen3.8-27B is a more recent alternative, not a reason to mix model-specific benchmark scores. Vendor benchmarks here are reference data, not Mac measurements.⁶⁵³

Source review October 10, 2026: Ollama package sizes and image modalities checked in the official tag list. Qwen’s benchmark scores were not independently reproduced on a Mac.

Sources

Footnotes

1. https://huggingface.co/Qwen/Qwen3.6-27B ↩ ↩²

2. https://huggingface.co/Qwen/Qwen3.6-35B-A3B ↩ ↩²

3. https://ollama.com/library/qwen3.6 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

4. https://ollama.com/library/qwen3.6/tags ↩ ↩² ↩³

5. https://qwen.ai/blog?id=qwen3.6-35b-a3b ↩ ↩²

6. https://qwen.ai/blog?id=qwen3.6-27b ↩ ↩²