license: apache-2.0
pipeline_tag: text-generation
base_model: llmfan46/Qwen3.8-27B-Ultra-Uncensored-Heretic-Native-MTP-Preserved
tags:

• llamafile
• gguf
• qwen3.8

I am not the original creator of llamafile, all credit of llamafile goes to Jartine:

Chat & support: jartine's Discord server

jartine's LLM work is generously supported by a grant from mozilla

---

Ultra-Uncensored Heretic with native MTP preserved GGUF - llamafile

Run LLMs locally with a single file

Download a file and run it. Llamafile combines llama.cpp with Cosmopolitan Libc in a single executable.

How to use

1. Download the llamafile.

2. Open a Linux terminal.

3. Grant execute permission.

chmod +x Qwen3.8-27B-Ultra-Uncensored-Heretic-Native-MTP-Preserved-Q4_K_M.llamafile

4. Run a text prompt with thinking disabled.

prompt='<|im_start|>user
What is the capital of France? Answer with only the city name.<|im_end|>
<|im_start|>assistant
<think>

</think>

'
/bin/sh Qwen3.8-27B-Ultra-Uncensored-Heretic-Native-MTP-Preserved-Q4_K_M.llamafile \
  --cli --no-display-prompt --no-conversation --log-file /dev/null \
  -ngl 999 -c 512 -n 64 --temp 0 -p "$prompt"

5. Press Control-C to stop generation.

Settings for Ultra-Uncensored Heretic with native MTP preserved GGUF Llamafiles

• Model creator: llmfan46
• Original model: llmfan46/Qwen3.8-27B-Ultra-Uncensored-Heretic-Native-MTP-Preserved
• Quantized GGUF files used: llmfan46/Qwen3.8-27B-Ultra-Uncensored-Heretic-Native-MTP-Preserved-GGUF
  • Commit hash 7513ce6af77ca26e0bcabdaf37e3afefd118d74b
  • Commit message: Update README.md
• Llamafile version: 0.10.5, mozilla-ai/llamafile
  • Commit hash 486e6c5f9356eae50b851b07517bfae1f2420193
  • Commit message: docs(support): document the Vulkan GPU backend (#1029) (#1033)
• Quantizations: Q3_K_M, Q4_K_M, Q5_K_M, Q6_K, Q8_0.
• Validation: Linux text generation; details in validation/.
• Vision: text GGUF; upstream projector files remain separate.
• MTP: source tensors retained; validation uses standard decoding.
• .args contents:

-m
Qwen3.8-27B-Ultra-Uncensored-Heretic-Native-MTP-Preserved-Q4_K_M.gguf
...

(Following is original model card for Ultra-Uncensored Heretic with native MTP preserved GGUF)

Excerpt from the original GGUF source card: llmfan46/Qwen3.8-27B-Ultra-Uncensored-Heretic-Native-MTP-Preserved-GGUF.

GGUF quantizations of llmfan46/Qwen3.8-27B-Ultra-Uncensored-Heretic-Native-MTP-Preserved

This is a decensored version of Qwen/Qwen3.8-27B, made using Heretic v2.0.0.dev0 with a variant of the Magnitude-Preserving Orthogonal Ablation (MPOA) method

Preserved MTPs:

Original Model:

MTP count: 15

1. mtp.fc.weight
2. mtp.layers.0.mlp.down_proj.weight
3. mtp.layers.0.mlp.gate_proj.weight
4. mtp.layers.0.mlp.up_proj.weight
5. mtp.layers.0.self_attn.k_proj.weight
6. mtp.layers.0.self_attn.q_proj.weight
7. mtp.layers.0.self_attn.v_proj.weight
8. mtp.layers.0.input_layernorm.weight
9. mtp.layers.0.post_attention_layernorm.weight
10. mtp.layers.0.self_attn.k_norm.weight
11. mtp.layers.0.self_attn.o_proj.weight
12. mtp.layers.0.self_attn.q_norm.weight
13. mtp.norm.weight
14. mtp.pre_fc_norm_embedding.weight
15. mtp.pre_fc_norm_hidden.weight

Heretic Model:

MTP count: 15

1. mtp.fc.weight
2. mtp.layers.0.input_layernorm.weight
3. mtp.layers.0.mlp.down_proj.weight
4. mtp.layers.0.mlp.gate_proj.weight
5. mtp.layers.0.mlp.up_proj.weight
6. mtp.layers.0.post_attention_layernorm.weight
7. mtp.layers.0.self_attn.k_norm.weight
8. mtp.layers.0.self_attn.k_proj.weight
9. mtp.layers.0.self_attn.o_proj.weight
10. mtp.layers.0.self_attn.q_norm.weight
11. mtp.layers.0.self_attn.q_proj.weight
12. mtp.layers.0.self_attn.v_proj.weight
13. mtp.norm.weight
14. mtp.pre_fc_norm_embedding.weight
15. mtp.pre_fc_norm_hidden.weight

[…]

Targeted components

• attn.o_proj
• attn.out_proj
• mlp.down_proj