The original BF16 weights are 55.6 GB. This conversion reduces the footprint to 14 GB, making it runnable on any Mac with 24 GB+ of unified memory with room to spare for large context windows.
Most local LLMs are "reactive" — they start generating a response before they've fully mapped out the logic. This model is deliberative.
Distilled from state-of-the-art Claude 4.6 Opus reasoning trajectories, it uses an advanced Chain-of-Thought (CoT) scaffold. Before providing its final answer, it enters an internal <think> state where it:
Deconstructs complex, multi-layered prompts into manageable sub-tasks
Simulates different solution paths and self-corrects logic errors before you see them
Reduces redundancy by adopting Claude's structured thinking pattern rather than the looping often seen in base reasoning models
This makes it the premier choice for technical planning, complex logic puzzles, and high-stakes decision support on Apple hardware.
Use apply_chat_template with enable_thinking=True. This is the idiomatic way to trigger the reasoning mode — no manual prompt construction needed.
from mlx_lm import load, generate
model, tokenizer = load("BeastCode/Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit")
messages = [
{
"role": "user",
"content": (
"A farmer needs to cross a river with a wolf, a goat, and a cabbage. "
"The boat can only hold the farmer and one other item. "
"If left alone, the wolf eats the goat, and the goat eats the cabbage. "
"How can he get everything across safely?"
),
}
]
# enable_thinking=True inserts the <think> prefix automatically
prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=True,
)
response = generate(model, tokenizer, prompt=prompt, max_tokens=2048, verbose=True)
print(response)
Tip: Set enable_thinking=False (or omit it) for fast, direct answers without the reasoning block — useful for simple lookups or chat where latency matters.
Manual prefill style (advanced)
If you want to see exactly what happens under the hood, you can construct the prompt manually and pre-fill the <think> token yourself. The two approaches produce identical results.
prompt = (
"<|im_start|>system\n"
"You are a highly analytical assistant.\n"
"<|im_end|>\n"
"<|im_start|>user\n"
"A farmer needs to cross a river with a wolf, a goat, and a cabbage. "
"The boat can only hold the farmer and one other item. "
"If left alone, the wolf eats the goat, and the goat eats the cabbage. "
"How can he get everything across safely?\n"
"<|im_end|>\n"
"<|im_start|>assistant\n"
"<think>\n"
)
response = generate(model, tokenizer, prompt=prompt, max_tokens=2048, verbose=True)
print(response)
The <think> block is invaluable for verifying logic, but you may want to strip it for a cleaner UI:
import re
def strip_thinking(text: str) -> str:
"""Remove the internal <think> process, returning only the final answer."""
return re.sub(r'<think>.*?</think>\s*', '', text, flags=re.DOTALL).strip()
mlx-community <huggingface.co/mlx-community>Newsgroups: alt.freeweights.qwenSubject: 
BAAI