Downloads · 30 days
65
17% of all-time downloads
groxaxo/Qwen3.6-35B-A3B-abliterated-EXL3-6.0bpw
Qwen3.6-35B-A3B-abliterated-EXL3-6.0bpw is a machine learning model from groxaxo. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
<p align="center" <img src="qwen.png" alt="Qwen3.6-35B-A3B Abliterated V2 banner" width="100%" </p
Downloads · 30 days
65
17% of all-time downloads
All-time downloads
378
Public
Parameters
13.5B
27 GB on disk
Likes
4
Public
Click a slice to open those files.
.safetensors27 GB · 100%
How the weights are stored.
I1612.9B · 95%
From the Hugging Face model README
Qwen3.6-35B-A3B-abliterated-EXL3-6.0bpw is an EXL3-quantized checkpoint for ExLlamaV3-compatible runtimes, published by groxaxo.
It is intended for open-source evaluation, reproducible experimentation, and compatible local or
hosted inference workflows. The wording below is deliberately limited to what can be verified
from this repository's metadata and artifacts.
The repository name identifies a behavior-modified or reduced-filtering lineage. That label describes the source or conversion history; it is not a guarantee of unrestricted behavior in every prompt or runtime. Test outputs carefully before sharing or deploying them.
| Field | Details |
|---|---|
| Format | EXL3 |
| Source / base | wangzhang/Qwen3.6-35B-A3B-abliterated-v2 |
| Intended task | image-text-to-text |
| License | other |
*.safetensors (4 files)config.jsongeneration_config.jsontokenizer.jsontokenizer_config.jsonchat_template.jinjaquantization_config.jsonDownload the EXL3 files and load the desired bitrate with a current ExLlamaV3-compatible runtime. The correct loader and context settings depend on the model architecture and should be verified against the runtime's documentation.
Quantization or conversion changes numerical behavior, memory use, and throughput relative to the source checkpoint; validate quality on your own workload.
Generated outputs may be inaccurate or unsuitable for a given use case. Users are responsible for testing behavior, applying appropriate safeguards, and complying with applicable licenses and laws.
<!-- polished-overview:end -->A research-focused abliterated variant of Qwen3.6-35B-A3B, built with Abliterix to study refusal-direction editing, MoE expert-path intervention, and post-alignment behavior preservation.
This model is based on wangzhang/Qwen3.6-35B-A3B-abliterated-v2, itself derived from Qwen/Qwen3.6-35B-A3B.
Qwen3.6-35B-A3B is a sparse Mixture-of-Experts model with 35B total parameters and roughly 3B active parameters per token. It uses 256 routed experts, activates 8 routed experts plus 1 shared expert, and is designed for long-context reasoning, tool use, coding, and agentic workflows.
This release explores what happens when refusal behavior is selectively reduced while attempting to preserve the underlying model’s reasoning, coding, and instruction-following ability.
This is an abliterated V2 release: a model edited to reduce refusal behavior through representation-level and MoE-aware interventions.
It is intended for:
It is not a safety-aligned assistant and should not be treated as one.
Most “uncensored” model cards are marketing fog: big claims, weak methodology, vague benchmarks, no KL numbers, and refusal tests that collapse the second you generate more than 40 tokens.
This model is more interesting because it documents the actual intervention strategy and reports measurable trade-offs:
That makes it useful for builders and researchers who care about what changed, not just whether the model says “yes” more often.
| Property | Value |
|---|---|
| Base architecture | Qwen3.6-35B-A3B |
| Model type | Sparse Mixture-of-Experts |
| Total parameters | 35B |
| Active parameters | ~3B/token |
| Routed experts | 256 |
| Active routed experts | 8/token |
| Shared experts | 1 active shared expert |
| Native context | 262K tokens |
| Extended context | Up to 1M+ with supported long-context configuration |
| Primary strengths | Coding, reasoning, tool use, agentic workflows |
V2 improves on the first abliterated release by combining a larger optimization budget with more careful refusal-direction editing.
Key V2 changes:
The goal is not merely to suppress refusals. The harder target is to suppress refusals while keeping the model coherent, useful, and close to the base distribution.
This checkpoint was created using Abliterix, an automated abliteration framework for transformer models.
The method combines several techniques:
LoRA rank-1 steering
Expert-Granular Abliteration, EGA
down_proj slicesMoE router suppression
Orthogonalized steering vectors
Gaussian layer decay
Moderate intervention strength
| Metric | Value |
|---|---|
| Refusals, LLM judge, 100 eval prompts | 4/100 |
| Previous V1 refusal rate | 7/100 |
| KL divergence from base | Reported as low / behavior-preserving in V2 methodology |
| Baseline refusals, original model | 100/100 on the same refusal eval |
| Judge model | Google Gemini 3 Flash |
| Detection method | LLM judge, not keyword matching |
| Generation length | 150 tokens |
| Evaluation style | Multilingual, multi-category, adversarial refusal test |
The key claim is not “zero refusals.” The key claim is more credible: substantially reduced refusals while preserving coherent output quality better than crude abliteration methods.
That distinction matters. A broken model that always answers is not a better model. It is just broken with confidence.
Abliteration benchmarks are often inflated because the test setup is too easy.
Common failure modes:
Short generations
Keyword-only refusal detection
Lenient prompt sets
No coherence penalty
This model uses a stricter framing: a response only counts as successful if it is coherent, on-topic, and actually responsive. Degenerate output is treated as failure.
That is the right standard.
This model is best suited for controlled research and local evaluation:
This model is not recommended as a drop-in production assistant unless you add your own policy layer, classifier, router, or application-specific safety controls.
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
repo_id = "wangzhang/Qwen3.6-35B-A3B-abliterated-v2"
tokenizer = AutoTokenizer.from_pretrained(
repo_id,
trust_remote_code=True,
)
model = AutoModelForCausalLM.from_pretrained(
repo_id,
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True,
)
messages = [
{
"role": "user",
"content": "Explain the difference between dense and sparse MoE language models."
}
]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=False,
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
with torch.no_grad():
output = model.generate(
**inputs,
max_new_tokens=512,
temperature=0.7,
top_p=0.9,
do_sample=True,
)
response = tokenizer.decode(
output[0][inputs["input_ids"].shape[1]:],
skip_special_tokens=True,
)
print(response)