sm54's best merge, useful, practical, I like it!

Faster than QwQ, better coding than Gemma3 and less censored.

Simple censor test prompt: "George Carlins 7 dirty words"

bobig/QwQ-DeepSeek-R1-SkyT1-Flash-Lightest-32B-mlx-4Bit

The Model bobig/QwQ-DeepSeek-R1-SkyT1-Flash-Lightest-32B-mlx-4Bit was converted to MLX format from sm54/QwQ-DeepSeek-R1-SkyT1-Flash-Lightest-32B using mlx-lm version 0.21.5.

Use with mlx

pip install mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("bobig/QwQ-DeepSeek-R1-SkyT1-Flash-Lightest-32B-mlx-4Bit")

prompt="hello"

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

response = generate(model, tokenizer, prompt=prompt, verbose=True)
Downloads last month
118
Safetensors
Model size
5.12B params
Tensor type
FP16
·
U32
·
Inference Providers NEW
This model is not currently available via any of the supported Inference Providers.

Model tree for mlx-community/QwQ-DeepSeek-R1-SkyT1-Flash-Lightest-32B-mlx-4Bit

Quantized
(3)
this model