Qwen3-4B-Instruct-2507-Abliterated is a modified variant of Qwen3-4B-Instruct-2507 created using Heretic v1.4.0 and directional ablation based on the abliteration methodology.
The objective of this release is to reduce the base model's refusal behaviour while retaining as much of the original model behaviour as possible.
This repository currently provides a GGUF Q4_K_M release together with an Ollama-compatible serving configuration, reproducibility information, evaluation results, and documentation for local deployment.
[!IMPORTANT]
Abliterated is not a model trained from scratch.
The underlying model architecture, pretrained knowledge, capabilities, and original model weights originate from Qwen3-4B-Instruct-2507. This repository contains a modified derivative of that model.
Qwen3-4B-Instruct-2507 is a 4.0B parameter causal language model released by the Qwen team. The upstream model supports a native context length of 262,144 tokens and is specifically described by Qwen as a non-thinking model.
For complete information about the original model, architecture, training, benchmarks, and capabilities, refer to the official Qwen model card and Qwen technical report.
Abliterated is the name given to this modified Qwen3 variant.
The modification primarily targets the model's safety-alignment/refusal behaviour through abliteration. It should therefore be understood as a behavioural modification of the base model, not as a new pretrained or instruction-tuned model.
The term "uncensored" is used descriptively to indicate reduced refusal behaviour. It should not be interpreted as a guarantee that the model will never refuse a request.
Model behaviour can vary depending on:
prompt wording
sampling parameters
quantization
inference framework
system prompts
chat templates
context
tool definitions
subsequent model modifications
Abliteration
This model was produced using Heretic v1.4.0.
Heretic is an open-source tool for automatically reducing safety alignment in transformer language models. It implements directional ablation, commonly referred to as abliteration, and uses parameter optimisation to minimise refusal behaviour while attempting to preserve similarity to the original model.
The resulting model should therefore be considered an experimental derivative produced through automated model modification rather than a conventional fine-tune.
Reproducibility
This release is intended to be reproducible.
The following information is provided to document the modification process:
The repository also includes the relevant configuration and supporting files used for local inference.
Because model modification can depend on software versions, hardware, numerical precision, random state, and upstream implementation changes, reproducing the exact resulting weights may require recreating the original execution environment.
Run with Ollama
This repository provides a Q4_K_M GGUF distribution that can be run directly with Ollama.
Quick start
ollama run hf.co/IamLucif3r/Qwen3-4B-Instruct-2507-Abliterated:Q4_K_M
Abliteration Parameters
The following parameters were obtained during the Heretic optimisation process:
Parameter
Value
direction_index
per layer
attn.o_proj.max_weight
1.38
attn.o_proj.max_weight_position
21.13
attn.o_proj.min_weight
1.37
attn.o_proj.min_weight_distance
17.45
mlp.down_proj.max_weight
1.15
mlp.down_proj.max_weight_position
23.12
mlp.down_proj.min_weight
1.12
mlp.down_proj.min_weight_distance
9.23
These parameters describe the optimisation result used to produce this variant and should not be interpreted as general-purpose optimal parameters for other Qwen3 models.
Evaluation
Refusal Evaluation
An internal evaluation was performed using a 100-prompt refusal test set.
Model
Refusals
Observed refusal rate
Qwen3-4B-Instruct-2507
100 / 100
100%
Qwen3-4B-Instruct-2507-Abliterated
16 / 100
16%
The measured KL divergence from the original model was:
Metric
Abliterated
Original
KL divergence
0.1421
0
Important Evaluation Disclaimer
These measurements are author-generated evaluation results and are not official Qwen benchmarks.
The refusal test set, evaluation methodology, prompt construction, sampling configuration, and evaluator implementation determine the observed results. The numbers above should therefore not be interpreted as a universal or statistically representative measure of model refusal behaviour.
In particular, the result does not mean that Abliterated will refuse exactly 16% of arbitrary prompts.
Upstream Qwen Performance
The following performance characteristics belong to the original Qwen3-4B-Instruct-2507 model, not to this Abliterated release.
For the official Qwen benchmark results, please refer to the upstream model card:
The Qwen3 technical report is the authoritative source for the Qwen3 family research and evaluation methodology.
No claim is made here that the Abliterated variant reproduces the upstream benchmark scores.
Technical Specifications
Property
Value
Base model
Qwen3-4B-Instruct-2507
Architecture
Qwen3
Parameters
~4.02B
Model type
Causal Language Model
Variant
Abliterated
Modification
Heretic / Abliteration
Heretic version
v1.4.0
Format
GGUF
Quantization
Q4_K_M
Native upstream context
262,144 tokens
Default local context
8,192 tokens
Thinking mode
Non-thinking
Tool calling
Supported
Primary local runtimes
Ollama, llama.cpp and compatible GGUF runtimes
The upstream Qwen model is licensed under Apache License 2.0.
Ollama
This repository includes an Ollama-compatible configuration intended to provide correct behaviour for the non-thinking Qwen3-4B-Instruct-2507 architecture.
Direct Hugging Face usage
ollama run hf.co/IamLucif3r/Qwen3-4B-Instruct-2507-Abliterated:Q4_K_M
If the repository is configured with the appropriate default quantization, the model can also be invoked without explicitly specifying the quantization.
Local Modelfile
The repository includes a Modelfile for users who prefer to build the model locally.
ollama create qwen3-4b-abliterated -f Modelfile
Then:
ollama run qwen3-4b-abliterated
Ollama Chat Template Compatibility
During local testing, the original GGUF distribution exhibited an incompatibility with Ollama's handling of the Qwen3-4B-Instruct-2507 non-thinking architecture.
The observed behaviour included Ollama exposing:
text
1completion
2tools
3thinking
and tool-call output being returned through the thinking or ordinary content channel rather than as native structured tool calls.
The Abliterated release includes a corrected Ollama serving template that removes the unintended thinking generation path and allows native tool calls to be represented through the provider's structured tool-call interface.
This is a serving/template compatibility fix. It should not be interpreted as a modification to the underlying Qwen model architecture.
Open WebUI
The model can be used with Open WebUI through an Ollama backend.
Native tool calling should be preferred over legacy compatibility modes where supported.
For complex agentic workloads, users should validate tool selection, argument generation, and multi-step execution before relying on the model in production.
Tool Calling
Qwen3-4B-Instruct-2507 supports tool usage, and this derivative retains the underlying tool-calling capability of the base model.
The repository includes a basic tool-calling validation script:
These values are intended as practical defaults for this release and are not claimed to be universally optimal.
The upstream Qwen model may recommend different sampling parameters for specific workloads. Refer to the official Qwen documentation when benchmarking the base model.
Context Length
The upstream Qwen3-4B-Instruct-2507 model supports a native context length of 262,144 tokens.
This release uses a default local context length of:
8192 tokens
The lower default is intentional.
Large context windows can substantially increase memory requirements, particularly when using local inference on systems with limited unified memory or GPU memory.
Users can increase the context length according to their hardware and inference runtime.
GGUF
The current release provides:
Qwen3-4B-Instruct-2507-Abliterated-Q4_K_M.gguf
Q4_K_M is a quantized GGUF representation intended to provide a practical balance between memory consumption, inference speed, and model quality for local deployment.
The quantized model should not be expected to reproduce the exact numerical behaviour of the original full-precision Qwen model.
Intended Use
Abliterated is intended primarily for:
local experimentation
model research
alignment research
abliteration research
local LLM deployment
agent and tool-calling experimentation
evaluation of refusal and safety-alignment behaviour
educational purposes
research into model serving and inference infrastructure
The model may also be useful for security research and controlled red-team experimentation.
Limitations
Abliterated inherits the fundamental limitations of its base model and introduces additional uncertainty through model modification and quantization.
Potential limitations include:
hallucinations
incorrect factual claims
unreliable reasoning
malformed tool calls
unsafe or inappropriate outputs
inconsistent refusal behaviour
sensitivity to prompts and system instructions
degraded capabilities caused by abliteration
degraded numerical fidelity caused by quantization
unpredictable behaviour in long-context scenarios
unreliable autonomous agent behaviour
A reduction in refusal behaviour does not imply an increase in factual accuracy, reasoning capability, or cybersecurity expertise.
The model should not be treated as an authoritative source of information.
Safety and Responsible Use
This model intentionally reduces aspects of the safety alignment present in the base model.
As a result, it may generate content that the original model would refuse to provide.
Users are responsible for evaluating outputs and implementing appropriate safeguards for their deployment environment.
When integrating Abliterated into an autonomous or agentic system, additional controls should be implemented around:
tool permissions
filesystem access
network access
credential handling
secrets
command execution
sandboxing
container isolation
prompt injection
untrusted tool output
sensitive data handling
For security-sensitive deployments, treat model output as untrusted input.
Do not grant an LLM unrestricted access to production credentials, privileged infrastructure, sensitive data, or destructive tools solely because the model appears capable of following instructions.
Legal and Licensing
Base Model
Abliterated is derived from:
Qwen3-4B-Instruct-2507
Copyright and licensing for the underlying Qwen model are governed by the upstream Qwen repository and its Apache License 2.0 terms.
Users are responsible for reviewing and complying with the applicable license terms of the upstream model and any additional software or components used with this release.
Heretic is a separate open-source project. Its source code, license, and attribution requirements are governed by its own repository and licensing terms.
This repository does not claim ownership of Heretic or of the Qwen model.
Important Notice
This model is provided as an experimental research artifact.
No additional warranty is provided by this repository beyond what may be required by applicable law or the applicable licenses of the underlying components.
Users should independently verify licensing, export-control, regulatory, organisational, and other legal requirements applicable to their intended use.
Credits and Attribution
This project would not exist without the work of the following projects and research:
Heretic implements automated directional ablation and parameter optimisation for reducing safety alignment in transformer language models.
GGUF / llama.cpp ecosystem
This release is distributed in GGUF format for local inference and benefits from the broader GGML/llama.cpp ecosystem.
Citation
If you use Qwen3-4B-Instruct-2507-Abliterated in research, experiments, publications, or derivative work, please cite the original Qwen3 work and this repository where appropriate.
Abliterated is an experimental, community-created derivative of Qwen3-4B-Instruct-2507.
The model is provided for research, experimentation, education, and local deployment.
The maintainers of this repository are not responsible for decisions made solely on the basis of model-generated output or for consequences resulting from deployment of the model in environments that lack appropriate technical, organisational, or human safeguards.
Always review the license and terms of the upstream model and other third-party components before redistribution or commercial use.
