These GGUF files are quantized using APEX, a novel MoE-aware mixed-precision quantization technique that outperforms standard quantization methods while being significantly smaller.

APEX beats Q8_0 perplexity at half the size — and even beats F16.

APEX classifies every tensor by its role — routed expert, shared expert, or attention — and applies a layer-wise precision gradient, giving the most sensitive edge layers higher precision and compressing the redundant middle layers more aggressively.

📦 APEX Quantization Tiers

File	Size	Profile	Best For
`*-APEX-I-Quality.gguf`	22 GB	I-Quality	Highest quality, best accuracy
`*-APEX-I-Balanced.gguf`	25 GB	I-Balanced	Best all-rounder, recommended
`*-APEX-I-Compact.gguf`	17 GB	I-Compact	Best quality/size ratio

📊 Why APEX?

Method	Size	Perplexity	HellaSwag	Speed
F16	64.6 GB	6.537	82.5%	30.4 t/s
Q8_0	34.4 GB	6.533	83.0%	52.5 t/s
APEX I-Quality	21.3 GB	6.552	83.5%	63.1 t/s
APEX I-Balanced	23.6 GB	6.548	83.0%	61.4 t/s
APEX I-Compact	16.1 GB	6.669	81.8%	69.8 t/s
APEX Mini	12.2 GB	7.088	81.0%	74.4 t/s

Benchmarks on Qwen3.5-35B-A3B, NVIDIA DGX Spark (GB10, 122 GB VRAM).

📐 I-Variant: Diverse Imatrix Calibration

I-variants use a diverse calibration dataset spanning chat, code, reasoning, and tool-calling — no Wikipedia. This produces higher accuracy on real-world benchmarks, lower KL divergence, and tiny perplexity increase on wikitext.

🔓 Original Model

Base: llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved

Decensored version of Qwen/Qwen3.6-35B-A3B, made using Heretic v1.3.0 + MPOA.

88% fewer refusals (10/100 Uncensored vs 83/100 Original) while preserving model quality (0.0015 KL divergence).

⚡ MTP Speculative Decoding

All 20 MTP layers preserved. Use --spec-type draft-mtp --spec-draft-n-max 2 for 20-50% speed boost.

🧠 Model Details

Architecture	Qwen3.5 MoE (Mixture of Experts)
Parameters	35B total, 3B active per token
Experts	256 routed experts, 8 active per token
Layers	40 transformer layers + 1 MTP layer
Context	262,144 tokens
Decensoring	Heretic v1.3.0 + MPOA (88% fewer refusals, 0.0015 KL)

🚀 Usage

llama.cpp

hf download SC117/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-APEX-GGUF --include "*.gguf" --local-dir ./models ./llama-server -m ./models/...APEX-I-Compact.gguf -ngl 99 -ncmoe 16 -c 131072 --spec-type draft-mtp --spec-draft-n-max 2

Ollama

echo 'FROM ./...APEX-I-Compact.gguf' > Modelfile ollama create my-model -f Modelfile && ollama run my-model

🎛️ Recommended Settings

Mode	Parameters
Thinking (general)	temp=1.0, top_p=0.95, top_k=20, pres_penalty=1.5
Thinking (coding)	temp=0.6, top_p=0.95, top_k=20
Instruct (general)	temp=0.7, top_p=0.8, top_k=20, pres_penalty=1.5
Instruct (reasoning)	temp=1.0, top_p=1.0, top_k=40, pres_penalty=2.0

Citation

bibtex

@misc{qwen36_35b_a3b,
title = {{Qwen3.6-35B-A3B}: Agentic Coding Power, Now Open to All},
url = {https://qwen.ai/blog?id=qwen3.6-35b-a3b},
author = {{Qwen Team}},
month = {April},
year = {2026}
}

cvgro/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-APEX-GGUF

Qwen3.6-35B-A3B

Links

Citation

cvgro/
Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-APEX-GGUF