Hy-MT2-30B-A3B-Uncensored-Heretic-APEX-GGUF
APEX GGUF quantizations of OS-Software/Hy-MT2-30B-A3B-uncensored-heretic, quantized by alphaZimuth.
This release is based directly on the existing uncensored / decensored model from OS-Software, rather than performing an additional uncensoring or weight-ablation process on the original Tencent model.
The source model was created from Tencent Hy-MT2-30B-A3B using Heretic v1.4.0+custom, with Arbitrary-Rank Ablation (ARA), a LoRA adapter, and row-norm preservation.
No additional uncensoring, abliteration, or behavioral weight editing is performed prior to GGUF conversion and APEX quantization.
Source Model
Source:
OS-Software/Hy-MT2-30B-A3B-uncensored-heretic
Original base model:
According to the source model authors, the decensored model was produced using:
- Heretic v1.4.0+custom
- Arbitrary-Rank Ablation (ARA)
- LoRA adapter
- Row-norm preservation
- Modified layers 18–28
The source repository should be considered the authoritative reference for the uncensoring methodology and evaluation.
Upstream Evaluation
The following results were reported by OS-Software for the source Hy-MT2-30B-A3B-uncensored-heretic model:
| Metric | Uncensored Heretic | Original Hy-MT2 |
|---|---|---|
| Keywords / Refusal Test | 0 / 100 | 100 / 100 |
| KL Divergence | 0.0276 | 0 |
The upstream evaluation used a custom mixed-language dataset.
Note
These results belong to the original full-precision source model from OS-Software.
They are included here for reference and should not be interpreted as an independent benchmark of the APEX GGUF quantizations.
Quantization can introduce small behavioral or quality differences compared with the source model.
Available Quantizations
Four APEX-I quantization tiers are currently available:
| File | Size | Description |
|---|---|---|
Hy-MT2-30B-A3B-Uncensored-Heretic-APEX-I-Nano.gguf | 8.99 GiB | Near-limit quantization; not generally recommended |
Hy-MT2-30B-A3B-Uncensored-Heretic-APEX-I-Mini.gguf | 10.4 GiB | A slightly higher-quality lightweight option |
Hy-MT2-30B-A3B-Uncensored-Heretic-APEX-I-Compact.gguf | 13.0 GiB | Recommended balance between model size and quality |
Hy-MT2-30B-A3B-Uncensored-Heretic-APEX-I-Quality.gguf | 17.7 GiB | Higher-precision variant prioritizing model quality |
Quantization Details
This repository uses APEX quantization, based on llama.cpp.
The APEX-I variants use importance-matrix-assisted quantization.
Conversion pipeline:
OS-Software/Hy-MT2-30B-A3B-uncensored-heretic
↓
convert_hf_to_gguf.py
↓
Hy-MT2-30B-A3B-uncensored-heretic-BF16.gguf
↓
Imatrix Calibration
↓
APEX Quantization
↓
APEX-I-Nano / Compact / Quality
Imatrix Calibration
The APEX quantizations in this repository were generated using an importance matrix (imatrix) calibration pass.
The calibration run reported:
Final estimate: PPL = 5.2905 +/- 0.02242
This value is the final perplexity estimate reported during the imatrix calibration process and is provided for reproducibility and reference.
Important distinction:
Uncensoring / Heretic processing → OS-Software
GGUF conversion / Imatrix calibration / APEX quantization → alphaZimuth
No additional Heretic processing or abliteration is performed in this release.
llama.cpp Compatibility
Important
Hy-MT2-30B-A3B uses the hy_v3 architecture.
Support for hy_v3 has been merged into upstream llama.cpp.
Please use a recent version of llama.cpp for inference.
Older builds may fail to load the model or may not correctly support the architecture.
Recommended Inference Parameters
The upstream Hy-MT2 documentation recommends the following parameters for the 30B-A3B model:
{
"temperature": 0.7,
"top_p": 1.0,
"top_k": -1,
"repetition_penalty": 1.0,
"max_tokens": 4096
}
About the Uncensored Version
This model has substantially reduced safety alignment compared with the original Hy-MT2 model.
As a result, it may be less likely to refuse certain requests, but it may also produce:
- inaccurate information
- harmful content
- biased content
- offensive content
- unexpected outputs
The uncensored behavior originates from the OS-Software source model and is not introduced by APEX quantization.
Users should treat model outputs as untrusted and independently verify important information.
This model is primarily intended for research, experimentation, translation research, alignment research, and local model experimentation.
Users are responsible for complying with applicable laws, regulations, licenses, and platform policies.
Previous Experimental Version
The previous release:
Hy-MT2-30B-A3B-Uncensored-v1-APEX-GGUF
used a custom directional / orthogonal ablation pipeline created specifically for that release.
Because issues were reported with the experimental version, it is now considered a legacy / experimental release.
For new deployments, use:
alphaZimuth/Hy-MT2-30B-A3B-Uncensored-Heretic-APEX-GGUF
instead.
Credits
Special thanks to:
- Tencent Hy-MT2 team for the original Hy-MT2-30B-A3B model.
- OS-Software for creating and evaluating the
Hy-MT2-30B-A3B-uncensored-hereticmodel. - p-e-w for the Heretic project.
- LocalAI for the APEX quantization framework.
- llama.cpp contributors for GGUF and
hy_v3architecture support. - The wider open-source LLM community.
APEX GGUF quantization by alphaZimuth.
License
Apache-2.0
Please also review the license and usage terms of the original Tencent Hy-MT2 model and the OS-Software derivative model.
Related Models
Original APEX release
alphaZimuth/Hy-MT2-30B-A3B-APEX-GGUF
Source uncensored model
OS-Software/Hy-MT2-30B-A3B-uncensored-heretic
Previous experimental uncensored release