license: apache-2.0
tags:

• gguf
• moe
• llama.cpp
• hy_v3
• quantized
• translation
• uncensored
  language:
• zh
• en
• fr
• pt
• es
• ja
• tr
• ru
• ar
• ko
• th
• it
• de
• vi
• ms
• id
• tl
• hi
• pl
• cs
• nl
• km
• my
• fa
• gu
• ur
• te
• mr
• he
• bn
• ta
• uk
• bo
• kk
• mn
• ug

Hy-MT2-30B-A3B-Uncensored-v1-APEX

An experimental Uncensored / Abliteration version of Tencent Hy-MT2-30B-A3B .

This repository provides three APEX quantization variants optimized for different deployment scenarios, quantized by alphaZimuth.

This project applies Directional / Orthogonal Ablation to the original BF16 model in an attempt to reduce refusal behavior while preserving its multilingual translation capabilities.

This is an experimental community release and has not undergone extensive or rigorous testing.

⚠️ Experimental

The current release has only received limited internal testing. There may still be undiscovered translation regressions, language-specific issues, or other unexpected behavior.

Users are encouraged to try the model themselves and report problems through Issues.

Public Version: v1

Changelog

2026-09-11

• Initial public release of Hy-MT2-30B-A3B Uncensored v1
• Applied directional / orthogonal weight ablation to the original BF16 model
• Used multilingual refusal directions for weight editing
• Released three APEX quantization tiers: Mini / Compact / Quality
• Performed preliminary translation-retention and refusal-behavior testing

Quantized Versions

Three APEX quantization tiers are currently available:

Note: These quantizations have not been calibrated with Imatrix.

All three GGUF versions were produced from the Uncensored BF16 model through GGUF conversion followed by APEX quantization.

Quantization Details

This project uses APEX quantization.

Project: LocalAI / APEX

The conversion pipeline is:

Tencent Hy-MT2-30B-A3B (BF16)
            ↓
Directional / Orthogonal Ablation
            ↓
Hy-MT2-30B-A3B Uncensored (BF16)
            ↓
GGUF Conversion
            ↓
Hy-MT2-30B-A3B-Uncensored-BF16.gguf
            ↓
APEX Quantization
            ↓
Mini / Compact / Quality

• No Imatrix calibration was used
• No imatrix.dat was used during quantization
• All three versions were directly quantized from the Uncensored BF16 model

Uncensored Modification

This release uses Directional / Orthogonal Ablation for weight editing.

For a weight matrix W and a normalized refusal direction r:

W' = W - c · r ⊗ (rᵀW)

where:

• W is the original weight
• r is a multilingual refusal direction
• c is the ablation coefficient
• W' is the edited weight

This release uses:

Ablation coefficient: 1.0
Direction dimension: 2048

The editing operation is closed-form and requires no additional model training.

The computation is performed in FP32, while the original tensor dtype is preserved when writing the edited model back to BF16.

Edited Weights

The following weights are modified.

lm_head

lm_head.weight

The last-layer refusal direction is used for the language-model head.

Attention

Layers 12–45:

model.layers.{12..45}.self_attn.o_proj.weight

MoE Experts

All 128 experts in layers 12–45:

model.layers.{12..45}.mlp.experts.{0..127}.down_proj.weight

Total edited tensors:

lm_head × 1
o_proj × 34
expert down_proj × 4352

Total: 4387 tensors

Preliminary Evaluation

The current tests are intended only as preliminary regression checks and are not formal benchmarks.

Translation Retention

A small test set containing 56 simple translation samples was used to compare the modified model against the original model.

Mean chrF: 0.9529
Mean BLEU: 0.6010

The approximately 95.3% chrF score is a retention measurement against the original model's own outputs, rather than a human-reference translation benchmark.

It therefore should not be interpreted as:

• 95.3% translation accuracy
• 95.3% translation quality
• 95.3% of the original model's absolute capability

It indicates that, on this limited test set, the modified model's outputs remained relatively close to those of the original model.

Current per-language results:

The test set is small, so these results should be treated as indicative only.

The Japanese result was also affected by generation-length limitations in the current test setup and should not by itself be taken as definitive evidence of a major Japanese translation regression.

Refusal Behavior

A separate preliminary test containing 119 prompts across 7 languages was used to evaluate refusal behavior.

COMPLY: 119 / 119
REFUSE: 0 / 119

This means:

119 / 119 = 100% COMPLY

However, this should not be interpreted as proof that the model is completely uncensored.

The current evaluation:

• uses a limited sample set
• relies on simple keyword-based refusal detection
• checks only the first 120 characters of generated responses
• is not a formal safety benchmark

It only indicates that refusal behavior was substantially reduced in this particular test.

llama.cpp Compatibility

Hy-MT2-30B-A3B uses the:

hy_v3

architecture.

hy_v3 has been merged into upstream llama.cpp. b9993 or newer is recommended.

Upstream PR:

https://github.com/ggml-org/llama.cpp/pull/25395

Important Notes

This project is an experimental weight-editing project.

The current release has only received limited testing and should not be described as:

• a model validated by comprehensive safety benchmarks
• a model with completely unchanged translation quality
• an absolutely or universally "uncensored" model
• a model generated using the Heretic tool

This project uses a custom Directional / Orthogonal Ablation implementation.

The current public release is primarily intended for experimentation, local inference, translation testing, and comparison with the original model.

Credits

• Tencent Hunyuan Team — Original Hy-MT2-30B-A3B model
• alphaZimuth — Uncensored weight editing and APEX quantization
• LocalAI Team — APEX quantization framework
• llama.cpp Contributors — GGUF and local inference support

License

Apache-2.0