---
language:
- en
- zh
license: apache-2.0
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- qwen3_5
- qwen3.8
- qwen
- merge
- nuslerp
- heretic
- abliterated
- uncensored
- cold-fusion
- fable
- vision-language
- mtp
- bf16
- conversational
base_model:
- gorbatjovy/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-heretic
- armand0e/Qwen3.8-27B-Fable-Distill-Heretic-ara
---

# Qwen3.8-27B-Cold-Fusion-Heretic-Fable-Heretic

A **NuSLERP merge of two abliterated Qwen3.8-27B derivatives**, combining the efficiency and reasoning-oriented behavior of **Cold-Fusion GAIN** with the creative / conversational characteristics of **Fable-Distill**.

This model uses the Heretic-abliterated versions of both source models:

- `gorbatjovy/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-heretic`
- `armand0e/Qwen3.8-27B-Fable-Distill-Heretic-ara`

The intent is to preserve the overall character of the existing Cold-Fusion + Fable merge recipe while using refusal-reduced upstream checkpoints.

## Merge recipe

The merge follows the same weighting pattern used by the Cold-Fusion/Fable NuSLERP recipe:

```yaml
models:
  - model: gorbatjovy/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-heretic
    parameters:
      weight: 1.6

  - model: armand0e/Qwen3.8-27B-Fable-Distill-Heretic-ara
    parameters:
      weight: 0.4

merge_method: nuslerp
dtype: bfloat16
```

For two-model NuSLERP, this corresponds to:

```text
t = 0.4 / (1.6 + 0.4)
  = 0.2
```

Merge parameters:

| Parameter | Value |
|---|---:|
| Cold-Fusion-Heretic weight | 1.6 |
| Fable-Heretic weight | 0.4 |
| NuSLERP `t` | 0.2 |
| Flatten | `true` |
| Row-wise | `false` |
| Output dtype | BF16 |

The merge was performed directly tensor-by-tensor using matching tensor names and shapes.

Both source checkpoints contain the same **1199 tensors**, with matching tensor keys and shapes.

This direct approach was used because current MergeKit architecture inference can incorrectly infer the hybrid attention layout of `Qwen3_5ForConditionalGeneration`.

## Source models

### Cold-Fusion-Heretic

`gorbatjovy/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-heretic`

This is an abliterated build of:

```text
Qwen/Qwen3.8-27B
        ↓
DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1
        ↓
Heretic abliteration
```

Cold-Fusion focuses on retaining strong general/reasoning capability while producing more concise reasoning behavior.

The Heretic version uses automated KL-constrained refusal-direction removal while retaining the Qwen3.8 vision-language tower and MTP head.

### Fable-Distill-Heretic-ARA

`armand0e/Qwen3.8-27B-Fable-Distill-Heretic-ara`

This is an abliterated build of the Fable-Distill branch:

```text
Qwen/Qwen3.8-27B
        ↓
Qwen3.8-27B-Fable-Distill
        ↓
Heretic v1.2.0 + ARA
```

The upstream model uses **Arbitrary-Rank Ablation (ARA)** with a rank-2 update over selected residual-writing projections.

Fable-Distill contributes more creative-writing, conversational, roleplay and narrative-oriented behavior to the merge.

## Architecture

The model retains the Qwen3.8 hybrid architecture:

- `Qwen3_5ForConditionalGeneration`
- 64 language-model layers
- hybrid linear attention + full attention
- native vision-language tower
- MTP speculative-decoding head
- BF16 safetensors
- 1199 tensors

The merge is performed over the full matching tensor set rather than only language-model layers.

## What this merge is intended to do

Very roughly:

```text
Cold-Fusion-Heretic
    ↓
reasoning / general capability
concise thinking behavior
task-oriented responses
            +
Fable-Distill-Heretic-ARA
    ↓
creative writing
natural dialogue
character voice
narrative style
            =
this merge
```

The Cold-Fusion branch remains dominant because of the `1.6 / 0.4` NuSLERP weighting.

This model is therefore intended as a fairly general-purpose Qwen3.8 derivative rather than a pure roleplay model.

## Abliteration notice

Both source checkpoints have undergone refusal-removal / abliteration.

This means the resulting merge may comply with prompts that the official aligned Qwen3.8 model would refuse.

Abliteration does **not** imply that the model is more accurate, reliable, or safe. It changes refusal behavior, not factual correctness.

Users are responsible for applying appropriate moderation and safeguards for their deployment environment.

## Formats / quantizations

### BF16

This repository contains the full BF16 merged checkpoint:

```text
CtrlCreeper/Qwen3.8-27B-Cold-Fusion-Heretic-Fable-Heretic
```

### MLX MXFP8

Apple Silicon / MLX version:

```text
CtrlCreeper/Qwen3.8-27B-Cold-Fusion-Heretic-Fable-Heretic-mxfp8-mlx
```

### MLX MXFP4

Lower-memory Apple Silicon / MLX version:

```text
CtrlCreeper/Qwen3.8-27B-Cold-Fusion-Heretic-Fable-Heretic-mxfp4-mlx
```

### GGUF

llama.cpp-compatible quantizations:

```text
CtrlCreeper/Qwen3.8-27B-Cold-Fusion-Heretic-Fable-Heretic-GGUF
```

Current / planned GGUF variants include:

- `Q4_K_M`
- `Q6_K`
- `Q8_0`

## Transformers usage

```python
import torch
from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "CtrlCreeper/Qwen3.8-27B-Cold-Fusion-Heretic-Fable-Heretic"

processor = AutoProcessor.from_pretrained(model_id)

model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {
        "role": "user",
        "content": "Explain why sqrt(2) is irrational."
    }
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
)

generated = outputs[0][inputs["input_ids"].shape[1]:]

print(
    processor.decode(
        generated,
        skip_special_tokens=True,
    )
)
```

## MLX usage

For example, with the MXFP8 build:

```bash
pip install -U mlx mlx-lm
```

```bash
mlx_lm.generate \
  --model CtrlCreeper/Qwen3.8-27B-Cold-Fusion-Heretic-Fable-Heretic-mxfp8-mlx \
  --prompt "Explain why sqrt(2) is irrational."
```

OpenAI-compatible server:

```bash
mlx_lm.server \
  --model CtrlCreeper/Qwen3.8-27B-Cold-Fusion-Heretic-Fable-Heretic-mxfp8-mlx
```

For systems with less unified memory, use the MXFP4 build instead.

## GGUF usage

Example with the `Q4_K_M` build:

```bash
llama-cli \
  -m Qwen3.8-27B-Cold-Fusion-Heretic-Fable-Heretic-Q4_K_M.gguf \
  -p "Explain why sqrt(2) is irrational." \
  -n 512
```

The GGUF files are converted from this BF16 Hugging Face checkpoint, not from the MLX quantizations.

## Merge implementation notes

The two source checkpoints were first verified to have:

```text
1199 tensors in Cold-Fusion-Heretic
1199 tensors in Fable-Distill-Heretic-ARA
0 missing tensor keys
0 extra tensor keys
0 tensor shape mismatches
```

NuSLERP was then applied tensor-by-tensor.

Floating-point tensors were interpolated in FP32 for numerical stability and converted back to the original BF16 dtype for storage.

Non-floating tensors, if present, use the dominant Cold-Fusion checkpoint value.

The output preserves the model configuration, tokenizer and processor metadata from the dominant Cold-Fusion-Heretic checkpoint.

## Credits

All model capability originates from the upstream models and their authors.

Special thanks to:

- Qwen team — `Qwen/Qwen3.8-27B`
- DavidAU — Cold-Fusion / GAIN work
- gorbatjovy — Cold-Fusion Heretic build
- Fable-Distill authors / maintainers
- armand0e — Fable-Distill Heretic ARA build
- Heretic contributors
- MergeKit contributors
- MLX / mlx-lm contributors
- llama.cpp contributors

This repository only provides the merged checkpoint and derived quantizations.

## Disclaimer

This model has reduced refusal behavior because both input checkpoints were intentionally abliterated.

Generated content may be incorrect, offensive, unsafe, misleading, or otherwise unsuitable for deployment without additional controls.

Do not assume compliance is equivalent to correctness.

Users are responsible for how the model is used and for complying with applicable laws, platform policies, and the inherited model license.

## License

Apache-2.0, inherited from the relevant upstream model lineage.