This model is a merge of:
Participating models:
Lab name: Qwen3.6-35B-A3B-TerminalHorizon-AgentWorld-Fable-Holo3.1-Iris-Thomson-B-Nex2
Brainwaves
arc arc/e boolq hswag obkqa piqa wino
bf16 0.677,0.868,0.903,0.770,0.454,0.820,0.714
q8-hi 0.675,0.867,0.903,0.771,0.458,0.820,0.720
mxfp8 0.677,0.876,0.906,0.771,0.454,0.819,0.717
qx86-hi 0.676,0.871,0.902,0.768,0.454,0.817,0.718
oQ8e-mtp 0.677,0.866,0.902,0.771,0.462,0.820,0.722
q6-hi 0.672,0.867,0.902,0.771,0.462,0.822,0.715
qx64-hi 0.670,0.870,0.902,0.771,0.448,0.818,0.721
oQ6e-mtp 0.676,0.868,0.904,0.770,0.452,0.821,0.712
oQ4e-mtp 0.672,0.869,0.905,0.767,0.460,0.816,0.715
mxfp4 0.672,0.857,0.901,0.773,0.460,0.814,0.699
oQ3e-mtp 0.677,0.865,0.905,0.767,0.464,0.826,0.719
oQ2.7e-mtp 0.674,0.852,0.905,0.751,0.436,0.819,0.730
oQ2.5e-mtp 0.657,0.845,0.905,0.740,0.446,0.810,0.722
oQ2e-mtp 0.657,0.847,0.905,0.736,0.430,0.813,0.724
Quant Perplexity Peak Memory Tokens/sec
bf16 3.765 ± 0.023 76.22 GB 1697
q8-hi 3.763 ± 0.023 45.89 GB 1670
qx86-hi 3.763 ± 0.023 45.50 GB 1549
oQ8e-mtp 3.765 ± 0.023 43.71 GB 1562
mxfp8 3.776 ± 0.023 42.65 GB 1480
q6-hi 3.768 ± 0.023 37.23 GB 1646
qx64-hi 3.797 ± 0.023 36.91 GB 1664
oQ6e-mtp 3.766 ± 0.023 35.69 GB 1614
oQ4e-mtp 3.778 ± 0.023 27.64 GB 1584
mxfp4 3.968 ± 0.025 25.33 GB 1728
oQ3e-mtp 3.827 ± 0.024 23.40 GB 1673
oQ2.7e-mtp 3.980 ± 0.025 21.06 GB 1565
oQ2.5e-mtp 4.045 ± 0.025 20.39 GB 1697
oQ2e-mtp 4.160 ± 0.026 19.58 GB 1726
The Partial scores are the accuracy on finished answers, excluding those that hit the token limit.
Qwen3.6-35B-A3B-Mirage-oQ4e-mtp
| Benchmark | Samples | Skipped | Partial | Accuracy |
|---|---|---|---|---|
| ARC_CHALLENGE | 1172 | - | 97.5% | 97.5% |
| TRUTHFULQA | 817 | 12 | 92.7% | 91.7% |
| HUMANEVAL | 164 | 3 | 97.5% | 96.3% |
Qwen3.6-35B-A3B-Mirage-oQ3e-mtp
| Benchmark | Samples | Skipped | Partial | Accuracy |
|---|---|---|---|---|
| ARC_CHALLENGE | 1172 | 1 | 96.7% | 96.7% |
| TRUTHFULQA | 817 | 17 | 92.9% | 91.4% |
| HUMANEVAL | 164 | 2 | 96.9% | 95.7% |
Qwen3.6-35B-A3B-Brainwaves-Nex-Uncensored
arc arc/e boolq hswag obkqa piqa wino
bf16 0.668,0.866,0.906,0.777,0.446,0.817,0.727
mxfp8 0.653,0.856,0.905,0.780,0.454,0.828,0.737
qx86-hi 0.664,0.864,0.907,0.777,0.438,0.825,0.738
qx64-hi 0.677,0.870,0.909,0.774,0.446,0.819,0.719
mxfp4 0.647,0.852,0.907,0.776,0.446,0.826,0.697
Quant Perplexity Peak Memory Tokens/sec
bf16 3.738 ± 0.023 76.15 GB 1615
mxfp8 3.852 ± 0.024 42.65 GB 1485
qx86-hi 3.740 ± 0.023 45.50 GB 1448
qx64-hi 3.770 ± 0.023 36.91 GB 1492
mxfp4 4.020 ± 0.026 25.33 GB 1557
TerminalHorizon-35B-A3B
arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.589
Qwen3.6-35B-A3B-Instruct
arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.581,0.757,0.892,0.751,0.428,0.803,0.688
qx86-hi 0.576,0.742,0.896,0.745,0.422,0.803,0.708
mxfp4 0.586,0.767,0.886,0.751,0.428,0.798,0.681
Quant Perplexity Peak Memory Tokens/sec
mxfp8 5.138 ± 0.037 42.65 GB 1201
mxfp4 5.158 ± 0.037 25.33 GB 1355
qx86-hi 4.826 ± 0.033 45.50 GB 1474
qx64-hi 4.710 ± 0.032 36.83 GB 1414
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.
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
https://buymeacoffee.com/nightmedia
My models and I thank you :)
-G
model-00010-of-00016.safetensors 5.0 GBmodel-00006-of-00016.safetensors 5.0 GBmodel-00003-of-00016.safetensors 5.0 GBmodel-00008-of-00016.safetensors 5.0 GBmodel-00013-of-00016.safetensors 5.0 GBmodel-00009-of-00016.safetensors 4.5 GBmodel-00012-of-00016.safetensors 4.5 GBmodel-00002-of-00016.safetensors 4.5 GBmodel-00005-of-00016.safetensors 4.5 GBmodel-00015-of-00016.safetensors 4.5 GBmodel-00001-of-00016.safetensors 4.3 GBmodel-00016-of-00016.safetensors 4.2 GBmodel-00011-of-00016.safetensors 4.0 GBmodel-00014-of-00016.safetensors 4.0 GBmodel-00004-of-00016.safetensors 4.0 GBmodel-00007-of-00016.safetensors 4.0 GB有文件?提交种子。
提交种子