leonsarmiento/Qwen3.6-27B-3bit-mlx

This model leonsarmiento/Qwen3.6-27B-3bit-mlx was converted to MLX format from Qwen/Qwen3.6-27B using mlx-lm version 0.31.2.

Quantization Details

The model uses mixed quantization:

• Embedding layers: 5-bit with group_size=64
• Prediction layers: 5-bit with group_size=64
• All other layers: 3-bit with group_size=64

This mixed precision approach provides a balance between compression and quality.

Use with mlx

pip install mlx-lm

Recommended Inference Parameters - use Jinja template on LM studio

Thinking Preserve ({%- set preserve_thinking = true %}):

• General tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Coding tasks: temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0

Instruct Mode ({%- set enable_thinking = false -%}):

• General tasks: temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Reasoning tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0

Example Usage

from mlx_lm import load, generate

model, tokenizer = load("leonsarmiento/Qwen3.6-27B-3bit-mlx")

prompt = "hello"

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

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

Example with Custom Parameters

from mlx_lm import load, generate

model, tokenizer = load("leonsarmiento/Qwen3.6-27B-3bit-mlx")

prompt = "hello"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_dict=False)

response = generate(
    model, 
    tokenizer, 
    prompt=prompt,
    temperature=0.6,
    top_p=0.95,
    top_k=20,
    min_p=0.0,
    repetition_penalty=1.0
)