SuperQwen-AgentWorld-35B-A3B-abliterated-mxfp4-mlx

Brainwaves

         arc   arc/e boolq hswag obkqa piqa  wino
mxfp4    0.646,0.838,0.902,0.778,0.444,0.822,0.703
Text only
mxfp4    0.657,0.862,0.906,0.766,0.490,0.825,0.692

Quant    Perplexity      Peak Memory   Tokens/sec
mxfp4    5.286 ± 0.038   25.33 GB

Base model

Qwen-AgentWorld-35B-A3B (VL)

         arc   arc/e boolq hswag obkqa piqa  wino
qx64-hi  0.644,0.818,0.909
mxfp4    0.626,0.813,0.901

Quant    Perplexity      Peak Memory   Tokens/sec
qx64-hi  3.954 ± 0.025   32.86 GB      1311
mxfp4    4.170 ± 0.028   25.33 GB      1599

Qwen-AgentWorld-35B-A3B-Text

         arc   arc/e boolq hswag obkqa piqa  wino
qx64-hi  0.647,0.835,0.909
mxfp4    0.626,0.813,0.901

Quant    Perplexity      Peak Memory   Tokens/sec
mxfp8    4.012 ± 0.026   42.65 GB      1543
qx64-hi  3.973 ± 0.026   32.86 GB      1532
mxfp4    4.170 ± 0.028   25.33 GB      1471

Thinking toggle

This model is using(an early version of) the fixed jinja template from froggeric/Qwen-Fixed-Chat-Templates

Drop <|think_on|> or <|think_off|> anywhere in your system or user prompt. The template intercepts the tag, removes it from context so the model never sees it, and flips the mode.

The tag syntax (<|think_on|>, <|think_off|>) uses Qwen's control-token delimiters, so it will never collide with real text. Earlier community templates used /think, which broke legitimate paths like cd /mnt/project/think.

I added a similar set of tags as <|think_forget|> or <|think_remember|> for handling the preserve_thinking flag.

-G

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("SuperQwen-AgentWorld-35B-A3B-abliterated-mxfp4-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)