Qwen3.5-122B-A10B-Text-qx64-hi-mlx

This model is Text only, the vision tower was removed.

Brainwaves

         arc   arc/e boolq hswag obkqa piqa  wino
qx85     0.456,0.519,0.622,0.704,0.392,0.774,0.680
qx64-hi  0.445,0.513,0.622,0.704,0.384,0.786,0.702
mxfp4    0.455,0.510,...

Quant    Perplexity     Speed(t/s)  Memory
qx85     3.762 ± 0.024  448         98.86 GB
qx64-hi  3.778 ± 0.024  435         98.28 GB
mxfp4    3.909 ± 0.025  535         71.94 GB

For comparison, the brainwaves from Qwen3-Coder-Next. The two architectures are structurally different.

         arc   arc/e boolq hswag obkqa piqa  wino
mxfp8    0.514,0.709,0.884,0.639,0.420,0.748,0.611
mxfp4    0.528,0.713,0.880,0.630,0.428,0.744,0.619
qx53n    0.520,0.714,0.872,0.630,0.438,0.744,0.599
qx64n-hi 0.527,0.707,0.880,0.631,0.426,0.744,0.580
qx64n    0.511,0.703,0.881,0.631,0.420,0.746,0.598
qx86n-hi 0.518,0.710,0.882,0.626,0.416,0.745,0.601
qx86n    0.515,0.712,0.881,0.627,0.414,0.744,0.590

The qx85 uses 5 bit data stores and 8 bit attention paths, embeddings, and head.

It was shaped to fit a 128GB Mac available RAM with a decent sized context.

The qx formula for the Qwen3.5 is still being refined, the model will be updated in-place if a more stable layer combination is found.

-G

This model Qwen3.5-122B-A10B-Text-qx64-hi-mlx was converted to MLX format from Qwen/Qwen3.5-122B-A10B using mlx-lm version 0.30.8.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("Qwen3.5-122B-A10B-Text-qx64-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)