Modelpublic

Ichigec/SuperQwen-AgentWorld-35B-A3B-abliterated-APEX-I-Quality

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes55downloads

Model Card

SuperQwen-AgentWorld-35B-A3B-abliterated — APEX-I-Quality GGUF

This is an APEX-I-Quality quantized GGUF of [SuperQwen-AgentWorld-35B-A3B-abliterated](https://huggingface.co/Jiunsong/SuperQwen-AgentWorld-35B-A3B-abliterated), a fused post-training checkpoint based on Qwen/Qwen-AgentWorld-35B-A3B.

What makes this model special

SuperQwen combines two post-training stages over the base Qwen-AgentWorld model:

1.Obliteratus false-refusal reduction — reduces unnecessary refusals on benign and authorized tasks via weight-space intervention
2.Supertune post-training — targeted tuning for AgentWorld observation formatting, direct task completion, JSON/tool formatting, Korean technical answers, and regression resistance

The APEX-I-Quality quantization (Adaptive Precision for EXpert Models) uses importance-matrix-guided mixed precision:

—Expert weights: Q6_K
—Shared weights: Q8_0
—Attention weights: Q6_K
—Imatrix calibration on 256K tokens of code/tools/math corpus
—Multi-Token Prediction (MTP) layers preserved at native precision

Why APEX v3 is better than Q8_0

Metric	SuperQwen Q8_0 (35 GB)	APEX v3 (21 GB)	Δ
Perplexity	5.837	5.870	+0.033 (within noise)
Size	35 GB	21 GB	−40% smaller
Speed (tg128)	—	40.2 t/s	—

The v3 achieves virtually identical quality at 40% smaller size compared to Q8_0, making it the best quality-to-size ratio for this model.

Full Benchmark Comparison

All benchmarks on identical hardware (DGX Spark, NVIDIA GB10, 99 GPU layers) using the APEX eval suite for comparability.

Metric	Qwen3.5 F16	Qwen3.5 Q8_0	Qwen3.5 APEX I-Q	Qwen3.6 Q8_0	Qwen3.6 APEX I-Q	Our Q8_0	Our APEX v1	Our APEX v3
Size	65 GB	34 GB	21 GB	~34 GB	21 GB	35 GB	22 GB	21 GB
PPL	6.537	6.533	6.552	6.720	6.735	5.837	5.868	5.870
HellaSwag	82.5%	83.0%	83.5%	82.5%	82.5%	—	82.75%	82.50%
Winogrande	74.5%	75.3%	74.5%	—	—	—	75.50%	75.50%
MMLU	41.5%	41.2%	41.4%	—	—	—	42.38%	41.93%
ARC	56.9%	57.9%	57.9%	—	—	—	54.52%	53.85%
tg128 (t/s)	30.4	52.5	63.1	9.7*	63.9	—	38.5	40.2

*Qwen3.6 Q8_0 tg128=9.7 — suspected different test configuration

Key takeaways

1.PPL: SuperQwen beats both base models — v3 PPL of 5.870 is 10% better than Qwen3.5 F16 (6.537) and 13% better than Qwen3.6 APEX I-Q (6.735). This is the supertune effect — agentic post-training improved text prediction accuracy.
2.HellaSwag: on par — within measurement noise of both base models at 82.5%.
3.MMLU: slight improvement — 41.9% vs 41.2-41.5% for Qwen3.5 base. Supertune added knowledge (MMLU-Pro: +14 points reported by Jiunsong).
4.ARC: slight regression — 53.9% vs 56.9-57.9% for Qwen3.5. Expected tradeoff from abliteration.

Quantization Details

—Method: APEX (Adaptive Precision for EXpert Models)
—Profile: i-quality (imatrix-guided, best accuracy)
—Base quant type: Q6_K
—Imatrix: 256,000 tokens of code/tools/math calibration corpus
—Layers: 40 (MoE with 256 experts, 8 per token)
—MTP: Preserved (1 layer, native precision)

Model Architecture

Parameter	Value
Architecture	Qwen3.5MoE
Total parameters	35B
Activated parameters	3B
Layers	40 (10 full attention + 30 linear attention)
Experts	256 (top-8 per token)
Context length	262,144 tokens
Vocabulary	248,320
MTP layers	1

Usage with llama.cpp

bash

# Download
hf download Ichigec/SuperQwen-AgentWorld-35B-A3B-abliterated-APEX-I-Quality \
  --local-dir ./models/

# Run
llama-cli \
  -m ./models/SuperQwen-APEX-I-Quality-v3.gguf \
  -ngl 99 \
  -c 32768 \
  -p "You are a helpful AI assistant."

File Information

File	Size	SHA256
`SuperQwen-APEX-I-Quality-v3.gguf`	21.3 GB	`afb6a47af5301e45d7e1de792e76d5611acb9d909b2a6d2a69fd645a12052162`

Credits

—Base model: Qwen/Qwen-AgentWorld-35B-A3B
—BF16 post-training: Jiunsong/SuperQwen-AgentWorld-35B-A3B-abliterated
—Quantization: APEX by mudler/apex-quant
—GGUF conversion & eval: llama.cpp + APEX eval suite
—Calibration corpus: code/tools/math (256K tokens)