Qwen3.6-35B-A3B-Qwable-Holo3-Qwopus
Supercalifragilisticexpialidocious (Mary Poppins, 1964)
supercaliflawjalisticexpialadoshus (The Daily Orange, 1931). The columnist muses about her made-up word, describing it as including "all words in the category of something wonderful" and "though rather long and tiring before one reaches its conclusion, ... once you arrive at the end, you have said in one word what it would ordinarily take four paragraphs to explain.
This is a NuSLERP merge of:
- llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved
- samuelcardillo/Qwopus-MoE-35B-A3B
- Hcompany/Holo3-35B-A3B
- lordx64/Qwable-v1
arc arc/e boolq hswag obkqa piqa wino
bf16 0.614,0.798,0.894,0.768,0.450,0.825,0.733
mxfp8 0.608,0.788,0.899,0.769,0.454,0.831,0.720
qx86-hi 0.610,0.796,0.896,0.768,0.452,0.826,0.735
qx64-hi 0.618,0.796,0.896,0.767,0.448,0.817,0.719
mxfp4 0.613,0.807,0.885,0.764,0.460,0.826,0.705
Quant Perplexity Peak Memory Tokens/sec
qx64-hi 4.118 ± 0.026 36.91 GB 1499
mxfp4 4.406 ± 0.028 25.33 GB 1414
Model components
Qwen3.6-35B-A3B-MTP-Holo3-Qwopus
arc arc/e boolq hswag obkqa piqa wino
bf16 0.603,0.774,0.895,0.756,0.428,0.808,0.713
mxfp8 0.608,0.770,0.897,0.761,0.430,0.814,0.707
qx86-hi 0.614,0.766,0.894,0.759,0.442,0.808,0.712
qx64y-hi 0.613,0.785,0.897,0.756,0.446,0.807,0.715
mxfp4 0.605,0.777,0.893,0.757,0.434,0.806,0.701
lordx64/Qwable-v1
arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.599,0.796,0.899,0.772,0.452,0.823,0.731
qx86-hi 0.614,0.800,0.900,0.770,0.458,0.820,0.732
qx64-hi 0.608,0.796,0.897,0.767,0.450,0.819,0.728
mxfp4 0.609,0.807,0.887,0.765,0.456,0.824,0.706
Baseline model
Qwen3.6-35B-A3B
arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.581,0.757,0.892,0.751,0.428,0.803,0.688
qx64-hi 0.589,0.763,0.892,0.748,0.428,0.801,0.700
mxfp4 0.586,0.767,0.886,0.751,0.428,0.798,0.681
This is not a Heretic model: the Heretic was used to restore MTP layers to Holo3-Qwopus before merging.
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.
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.
- The template intercepts the tag, removes it from context so the model never sees it, and flips the mode.
The model has not been tested with the latest version of froggeric's jinja template, and it is possible that the performance would go up with the latest version, I just did not have enough time to test it.
-G
Model recipe
models:
- model: lordx64/Qwable-v1
parameters:
weight: 1.6
- model: Qwen3.6-35B-A3B-MTP-Holo3-Qwopus
parameters:
weight: 0.4
merge_method: nuslerp
dtype: bfloat16
name: Qwen3.6-35B-A3B-Qwable-Holo3-Qwopus