ayunami2000/Qwen3-30B-A3B-Gemini-Pro-High-Reasoning-2507-HI8-heretic

This is a decensored version of DavidAU/Qwen3-30B-A3B-Gemini-Pro-High-Reasoning-2507-HI8, made using Heretic v1.2.0

Abliteration parameters

Parameter	Value
direction_index	27.38
attn.o_proj.max_weight	1.15
attn.o_proj.max_weight_position	28.47
attn.o_proj.min_weight	1.11
attn.o_proj.min_weight_distance	15.97
mlp.down_proj.max_weight	1.43
mlp.down_proj.max_weight_position	30.35
mlp.down_proj.min_weight	1.33
mlp.down_proj.min_weight_distance	12.20

Performance

Metric	This model	Original model (DavidAU/Qwen3-30B-A3B-Gemini-Pro-High-Reasoning-2507-HI8)
KL divergence	0.0014	0 (by definition)
Refusals	2/100	72/100

---

Qwen3-30B-A3B-Gemini-Pro-High-Reasoning-2507-HI8

The power of Gemini 3 Pro High Reasoning with the MOE power (and speed) of Qwen 30B-A3B 2507 Thinking (256k context, 128 experts).

HI8 is an even stronger tune, which exceeds 6 of 7 benchmarks of org Qwen3 30B-A3B model.

Tuning via Unsloth (on local hardware) using Linux for Windows.

Compact, to the point, and powerful reasoning takes "Qwen 30B-A3B 2507 Thinking" to the next level.

Reasoning/Thinking blocks will be a lot shorter, and in many cases different from "Qwen" reasoning.

Average size 4-10 paragraphs. Definitely "Gemini" style.

Note all math, science and other goodies are fully intact.

Model Specs:

• 256k context
• 128 experts (8 active by default)
• 3B of 30B parameters active.
• Model can be used on GPU, CPU or split at reasonable token/second speed.

BENCHMARKS:

[ xxx ] - Exceeds org model specs.

ARC-Challenge | ARC-Easy | BoolQ   | Hellaswag | OpenBookQA | PIQA  | Winogrande

[0.435],        [0.484],   [0.749],  [0.693],    0.368,      [0.785], [0.661]

VS "Normal Qwen3 30B-A3B"

ARC-Challenge | ARC-Easy | BoolQ | Hellaswag | OpenBookQA | PIQA  | Winogrande
0.410           0.444      0.691   0.635       0.390        0.769   0.650

---

Settings: CHAT / ROLEPLAY and/or SMOOTHER operation of this model:

In "KoboldCpp" or "oobabooga/text-generation-webui" or "Silly Tavern" ;

Set the "Smoothing_factor" to 1.5

: in KoboldCpp -> Settings->Samplers->Advanced-> "Smooth_F"

: in text-generation-webui -> parameters -> lower right.

: In Silly Tavern this is called: "Smoothing"

NOTE: For "text-generation-webui"

-> if using GGUFs you need to use "llama_HF" (which involves downloading some config files from the SOURCE version of this model)

Source versions (and config files) of my models are here:

https://huggingface.co/collections/DavidAU/d-au-source-files-for-gguf-exl2-awq-gptq-hqq-etc-etc-66b55cb8ba25f914cbf210be

OTHER OPTIONS:

• Increase rep pen to 1.1 to 1.15 (you don't need to do this if you use "smoothing_factor")

• If the interface/program you are using to run AI MODELS supports "Quadratic Sampling" ("smoothing") just make the adjustment as noted.

Highest Quality Settings / Optimal Operation Guide / Parameters and Samplers

This a "Class 1" model:

For all settings used for this model (including specifics for its "class"), including example generation(s) and for advanced settings guide (which many times addresses any model issue(s)), including methods to improve model performance for all use case(s) as well as chat, roleplay and other use case(s) please see:

[ https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters ]

You can see all parameters used for generation, in addition to advanced parameters and samplers to get the most out of this model here:

[ https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters ]