         arc   arc/e boolq hswag obkqa piqa  wino
qx86-hi  0.420,0.457,0.379,0.671,0.354,0.777,0.702
qx64-hi  0.413,0.459,0.378,0.670,0.366,0.772,0.687
mxfp4    0.413,0.464,0.378,0.675,0.364,0.771,0.687

Quant    Perplexity     Max Memory
qx86-hi  4.042 ± 0.026  41.52 GB
qx64-hi  4.073 ± 0.026  32.86 GB

Older models with VL

Qwen3-VL-30B-A3B-Instruct
qx86-hi  0.439,0.541,0.894,0.619,0.430,0.764,0.592
qx64-hi  0.454,0.544,0.893,0.618,0.428,0.749,0.590

Qwen3-VL-30B-A3B-Thinking
qx86-hi  0.393,0.466,0.751,0.648,0.366,0.776,0.667

For a thinking model, it performs above the previous VL MoE.

Deckard formula update

This quant has been updated with the new Deckard(qx) formula for Qwen3.5, and includes some attention layers new to the architecture.

While the changes seem minimal for the added weight, the model is more confident, with better attention to detail, and the vibe quality has improved.

Old formula metrics

qx86-hi  0.418,...
qx64-hi  0.419,...

Perplexity
qx86-hi  4.037 ± 0.026   37.46 GB
qx64-hi  4.079 ± 0.026   28.79 GB

More metrics will be available soon

-G

This model Qwen3.5-35B-A3B-Text-qx86-hi-mlx was converted to MLX format from Qwen/Qwen3.5-35B-A3B using mlx-lm version 0.30.8.

Use with mlx

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("Qwen3.5-35B-A3B-Text-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, return_dict=False,
    )

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

nightmedia/Qwen3.5-35B-A3B-Text-qx86-hi-mlx

Qwen3.5-35B-A3B-Text-qx86-hi-mlx

Deckard formula update

Use with mlx

nightmedia/
Qwen3.5-35B-A3B-Text-qx86-hi-mlx