---
language:
  - en
  - zh
license: apache-2.0
library_name: gguf
pipeline_tag: text-generation
base_model: AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
base_model_relation: quantized
quantized_by: vcruz305
tags:
  - gguf
  - qwen
  - qwen3.8
  - llama.cpp
  - uncensored
  - abliterated
  - aeon
  - mtp
---

# Qwen3.8-27B AEON Ultimate Uncensored GGUF

llama.cpp K-quants of [AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16](https://huggingface.co/AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16), an abliterated BF16 of [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B).

This is **not** official Qwen. It is also **not** [vcruz305/Qwen3.8-27B-GGUF](https://huggingface.co/vcruz305/Qwen3.8-27B-GGUF) (official trunk) or [vcruz305/Qwen3.8-27B-Uncensored-GGUF](https://huggingface.co/vcruz305/Qwen3.8-27B-Uncensored-GGUF) (orcarouter FP8).

**MTP is baked into every Q2–Q8 file** (866 tensors, `qwen35.nextn_predict_layers=1`, +~0.24 GiB vs the old trunk-only files). You do **not** need a second GGUF for `draft-mtp`.

## What is in these files

27B **dense** hybrid-attention (`qwen35`). **64** language-trunk blocks plus **1** nextn/MTP block. Converted from the AEON BF16 master (no `--no-mtp`). Pair a separate `mmproj` if you need vision.

The source removed the refusal direction. These GGUFs inherit that.

## Chat template

Official 3.8 jinja wraps empty `<think>` blocks and breaks multi-turn agents. **These GGUFs bake a fixed template.** Use `--jinja`.

```bash
llama-server -m Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-Q4_K_M.gguf --jinja --reasoning-format deepseek
```

## Files

| File | Quant | Bytes | GiB | Notes |
|---|---|---:|---:|---|
| `Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-Q2_K.gguf` | Q2_K | 10864592928 | 10.12 | live; MTP baked in |
| `Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-Q3_K_M.gguf` | Q3_K_M | 13500737568 | 12.57 | live; MTP baked in |
| `Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-Q4_K_M.gguf` | Q4_K_M | 16810715168 | 15.66 | live; MTP baked in |
| `Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-Q5_K_M.gguf` | Q5_K_M | 19535702048 | 18.20 | live; MTP baked in |
| `Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-Q6_K.gguf` | Q6_K | 22431000608 | 20.89 | live; MTP baked in; largest full-GPU on 24GB Turing |
| `Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-Q8_0.gguf` | Q8_0 | 29047085088 | 27.05 | live; MTP baked in; will not `-ngl 99` on 24GB |

## How to run MTP

Needs a llama.cpp build that understands Qwen3.5 nextn / `draft-mtp`. `--parallel 1` is required.

```bash
llama-server \
  -m Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-Q4_K_M.gguf \
  --spec-type draft-mtp --spec-draft-n-max 2 --spec-draft-p-min 0.7 \
  --parallel 1 --jinja --reasoning-format deepseek \
  -a qwen38-27b-aeon --host 127.0.0.1 --port 8085 \
  -ngl 99 -fa on -b 512 -ub 512 -c 32768
```

Optional extra `mtp-*` sidecars are still in the repo if you want a separate `-md` draft. They are **not** required for the files above.

## Download

Use hf_xet. Do not `git clone`. Grab only the file that exists.

```bash
export HF_XET_HIGH_PERFORMANCE=1
hf download vcruz305/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-GGUF \
  --local-dir Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-GGUF \
  --include "Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-Q4_K_M.gguf"
```

Q6_K is the largest file that still full-offloads a 24GB Turing card. Q8_0 does not.

## Intended use

Local llama.cpp of the AEON uncensored 27B trunk + native MTP. Out of scope: treating this as official Qwen, or as a drop-in for the official-trunk GGUF pack.

## Source

- BF16 master: https://huggingface.co/AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
- Official base: https://huggingface.co/Qwen/Qwen3.8-27B
- Convert: `convert_hf_to_gguf.py --outtype f16` (MTP mixin on, no `--no-mtp`) → `llama-quantize`
- License: Apache-2.0

## Credits

Abliteration / BF16: [AEON-7](https://huggingface.co/AEON-7). Base: Qwen / Alibaba. GGUF pack: Victor Cruz ([vcruz305](https://huggingface.co/vcruz305)).