KAT-Dev-72B-Exp-qx86-hi-mlx

This is the Deckard(qx) quant formula--a tuned mixed precision design.

On regular Qwens it makes the model "happier", and more involved in the interaction, often using metaphors to anchor meaning(see other model cards for examples)

I have no idea how well it works on this one, running metrics on this will take a while.

If you "like" the model, I'll keep it in the collection.

Model           Perplexity  Peak memory
qx86-hi-mlx  4.333 ± 0.031     71.47 GB
q8-hi        4.334 ± 0.031     86.59 GB

The mixed precision model is slightly outperforming q8. This is expected

-G

This model KAT-Dev-72B-Exp-qx86-hi-mlx was converted to MLX format from Kwaipilot/KAT-Dev-72B-Exp using mlx-lm version 0.28.2.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("KAT-Dev-72B-Exp-qx86-hi-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
    )

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