Architecture note: Qwen3.6-27B uses the qwen3_5 hybrid architecture — a mix of Gated DeltaNet (linear_attention) and full GQA layers in a 3:1 ratio across 64 layers. This is distinct from the standard Qwen3 transformer.

Use with mlx

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/Qwen3.6-27B-4bit")

prompt = "hello"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True
    )

response = generate(model, tokenizer, prompt=prompt, verbose=True)

Thinking mode

Qwen3.6 supports a thinking mode for complex reasoning. Enable it by including /think in your prompt (or using the chat template's thinking flag):

python

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/Qwen3.6-27B-4bit")

messages = [{"role": "user", "content": "Think step by step: what is 17 × 23?"}]
prompt = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    enable_thinking=True,
)

response = generate(model, tokenizer, prompt=prompt, max_tokens=1024, verbose=True)

Hardware requirements

Precision	Approx. Size	Min Unified RAM
4-bit (this model)	~14 GB	32 GB
8-bit	~28 GB	64 GB

Benchmarks

Tested on Apple M5 Pro (64 GB) with macOS 26.4, mlx-lm 0.31.3.

Metric	4-bit	8-bit
Load time	1.8 s	5.6 s
Prompt throughput	16.5 tok/s	5.7 tok/s
Generation throughput	18.3 tok/s	9.4 tok/s
Peak memory	15.4 GB	30.5 GB

raven88/Qwen3.6-27B-4bit

mlx-community/Qwen3.6-27B-4bit

Use with mlx

Thinking mode

Hardware requirements

Benchmarks

raven88/
Qwen3.6-27B-4bit