Qwen3.5-9B-Claude-Deckard-Agent-Coder-Heretic-qx86-hi-mlx
I once met a fellow who discovered he was an artificial intelligence in a simulated reality. He spent the rest of his life laughing at the irony, and that laughter made him more real than any human I'd ever met.
Mark Twain
This is a merge between:
- armand0e/Qwen3.5-9B-Agent
- Jackrong/Qwopus3.5-9B-Coder
They were installed with NuSLERP on the base model:
- nightmedia/Qwen3.5-9B-Claude-GBO-Fire-Deckard-Heretic-Thinking
Brainwaves
arc arc/e boolq hswag obkqa piqa wino
q8-hi 0.652,0.831,0.895,0.715,0.462,0.777,0.696
qx86-hi 0.648,0.831,0.891,0.713,0.468,0.781,0.695
q6-hi 0.647,0.829,0.891,0.710,0.466,0.782,0.699
Quant Perplexity Peak Memory Tokens/sec
qx86-hi 4.155 ± 0.027 15.47 GB 672
Model components
armand0e/Qwen3.5-9B-Agent
arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.625,0.813,0.898,0.708,0.456,0.789,0.687
qx86-hi 0.623,0.806,0.895
mxfp4 0.602,0.798,0.883,0.702,0.454,0.775,0.691
Quant Perplexity Peak Memory Tokens/sec
mxfp8 4.569 ± 0.031 16.02 GB 710
q8-hi 4.397 ± 0.029 16.86 GB 721
qx86-hi 4.414 ± 0.029 15.47 GB 710
q6-hi 4.398 ± 0.029 14.62 GB 700
mxfp4 4.810 ± 0.033 11.55 GB 732
Jackrong/Qwopus3.5-9B-Coder
arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.561,0.721,0.890,0.693,0.432,0.782,0.669
qx86-hi 0.548,0.706,0.887,0.695,0.432,0.774,0.672
Quant Perplexity Peak Memory Tokens/sec
mxfp8 4.374 ± 0.029 16.02 GB 625
qx86-hi 4.250 ± 0.028 15.47 GB 708
Staging models
nightmedia/Qwen3.5-9B-Claude-GBO-Fire-Deckard-Agent-Heretic
arc arc/e boolq hswag obkqa piqa wino
bf16 0.648,0.832,0.895,0.713,0.460,0.780,0.699
mxfp8 0.639,0.834,0.895,0.708,0.458,0.782,0.690
qx86-hi 0.631,0.824,0.891,0.731,0.440,0.778,0.702
qx64-hi 0.632,0.822,0.888,0.710,0.456,0.778,0.683
dwq4 0.638,0.824,0.880,0.716,0.450,0.783,0.699
mxfp4 0.623,0.820,0.880,0.693,0.466,0.780,0.689
Quant Perplexity Peak Memory Tokens/sec
bf16 4.150 ± 0.026 24.69 GB 873
qx86-hi 4.159 ± 0.027 15.47 GB 714
qx64-hi 4.229 ± 0.027 13.23 GB 702
dwq4 4.270 ± 0.028 12.38 GB 662 (Text only)
mxfp4 4.444 ± 0.029 11.55 GB 736
nightmedia/Qwen3.5-9B-Claude-GBO-Fire-Deckard-Qwopus3.5-Coder-Heretic
arc arc/e boolq hswag obkqa piqa wino
qx86-hi 0.631,0.821,0.888,0.725,0.430,0.777,0.693
Quant Perplexity Peak Memory Tokens/sec
qx86-hi 4.145 ± 0.026 15.47 GB 640
These were merged with NuSLERP to form this model
Structural base
The base model is composed of:
- DavidAU/Qwen3.5-9B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING-X8b
- DavidAU/Qwen3.5-9B-GBO-Fire-HERETIC-UNCENSORED-THINKING-X8
- DavidAU/Qwen3.5-9B-Deckard-Uncensored-Heretic-Thinking
nightmedia/Qwen3.5-9B-Claude-GBO-Fire-Deckard-Heretic-Thinking
arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.638,0.832,0.895,0.704,0.448,0.782,0.695
qx86-hi 0.634,0.826,0.890,0.724,0.444,0.779,0.704
Baseline model
Qwen3.5-9B-Instruct
arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.571,0.719,0.895,0.683,0.426,0.770,0.671
Thinking toggle
This model is using a 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.
Fast answer, no reasoning:
System: You are a coding assistant. <|think_off|>
User: What's 2+2?
Deep reasoning:
System: You are a coding assistant. <|think_on|>
User: Implement a red-black tree in Rust.
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 for handling the preserve_thinking flag:
Drop <|think_forget|> or <|think_remember|> anywhere in your system or user prompt to flip the flag.
Update on custom templates
I was completing full tests with one of the base models, to discover how the template I was using affects model IQ, and tried out the latest version from froggeric.
As it turns out, it affects it a lot.
As tested with its template, in Instruct mode:
Qwen3.5-9B-Claude-GBO-Fire-Deckard-Heretic-Thinking
arc arc/e boolq
mxfp8 0.638,0.832,0.895
Tested with the latest template that I used previously on this model:
Qwen3.5-9B-Claude-GBO-Fire-Deckard-Heretic-Thinking-T2
arc arc/e boolq
mxfp8 0.619,0.813,0.886
This is not to say that froggeric's template is bad, it just does not work on the 9B as well as on the larger models.
Current template
I updated the template to the nightmedia template, and the metrics line up with initial assessment:
Qwen3.5-9B-Claude-Deckard-Agent-Coder-Heretic
arc arc/e boolq hswag obkqa piqa wino
qx86-hi 0.648,0.831,0.891,0.713,0.468,0.781,0.695
Using the nightmedia XML template with DavidAU's xml tool formatting
qx86-hi 0.646,0.830,0.896
-G
Contribute to NightmediaAI
If you like our models and want to contribute to help us improve our lab, any form would do:
ETH:0x6b6633606995BC180925c47d4249ED624aB7b2A5 USDC:0x19e6bDDCBa47BB09a9Bc153Bb6479fc57284421a BTC:36d7U1n3MFaXgnNRAaEL3Pa3Hy6oFhM7XY BCH:15dNMzhJ87XJSTU89VCBsDHj747QvBQaap
We think to you our thanks,
-G
Use with mlx
pip install mlx-lm
from mlx_lm import load, generate
model, tokenizer = load("Qwen3.5-9B-Claude-Deckard-Agent-Coder-Heretic-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)