license: cc-by-nc-4.0
base_model:
- JohnDi/Hemmingway-1-Abliterated-Extreme-GGUF
- Altworld/Hemmingway-1
tags: - gguf
- qwen3.8
- qwen35
- abliterated
- uncensored
- mtp
- llama.cpp
pipeline_tag: text-generation
Hemmingway-1 Abliterated Extreme Q8_0 + MTP
A single-file Q8_0 GGUF of JohnDi's abliterated Hemmingway-1 Extreme, with the complete MTP decoder block from the original Altworld Hemmingway-1 Q8_0 GGUF added for llama.cpp speculative decoding.
This is an unofficial community derivative. It is not published or endorsed by Altworld, JohnDi, or bartowski.
File
Hemmingway-1-Abliterated-Extreme-Q8_0-MTP.gguf— 29.05 GB (Q8_0)
The abliterated model's 64-layer trunk and chat template are retained. The donor's 15-tensor MTP block and required GGUF metadata are added as block 64. This makes the resulting GGUF declare 65 blocks, including one MTP prediction layer. MTP tensors are the original non-abliterated model's weights; they were not retrained against the abliterated trunk.
Run with llama.cpp
Use a recent llama.cpp build with Qwen3.5 MTP support:
llama-server \
--model Hemmingway-1-Abliterated-Extreme-Q8_0-MTP.gguf \
--ctx-size 163840 \
--n-gpu-layers 999 \
--spec-type draft-mtp \
--spec-draft-n-max 4 \
--reasoning on \
--reasoning-format deepseek-legacy \
--jinja
Hardware limits may require a lower context size or partial GPU offload. The GGUF includes the model's chat template. MTP loaded and generated successfully with llama.cpp on ROCm. One local short-use sample accepted 787 of 1,992 proposed draft tokens; this is not a speed benchmark, and acceptance varies with prompt and hardware.
Lineage and attribution
- Base model: Altworld/Hemmingway-1, based on Qwen/Qwen3.8-27B.
- Abliterated GGUF: JohnDi/Hemmingway-1-Abliterated-Extreme-GGUF.
- MTP donor GGUF: bartowski/Altworld_Hemmingway-1-GGUF, Q8_0.
- The donor's MTP block was copied into the abliterated GGUF; the base model weights were not merged or re-quantized.
License
The Altworld model card currently identifies Hemmingway-1 as CC BY-NC 4.0 and permits sharing with attribution for non-commercial use. Some upstream GGUF cards label their files Apache-2.0, so upstream license labels conflict. This repository uses CC BY-NC 4.0 to honor the more restrictive base-model terms; it does not grant commercial-use rights. Check with the respective rights holders if you need a definitive interpretation or commercial permission.
Credit Altworld for Hemmingway-1, JohnDi for the abliterated derivative, and bartowski for the Q8_0 MTP donor conversion.