license: apache-2.0
license_link: LICENSE
base_model:
- HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
- unsloth/Qwen3.6-35B-A3B-MTP-GGUF
base_model_relation: merge
tags: - uncensored
- qwen3.6
- moe
- gguf
- mtp
- speculative-decoding
pipeline_tag: image-text-to-text
Qwen3.6-35B-A3B Uncensored (HauhauCS Aggressive) Q4_K_M, with MTP head
This is an uncensored model. Its refusals were trained out by its author, so it
answers what the base model would decline. You are responsible for how you use it.
HauhauCS's uncensored Qwen3.6-35B-A3B at Q4_K_M, with the multi-token-prediction
(MTP) head from unsloth's conversion of the original model grafted on, so
llama.cpp can draft with it (--spec-type draft-mtp). This is the file
mecha installs as the catalog modelqwen3.6-35b-a3b-uncensored; hosting it saves building the graft locally.
What was changed
Nothing was retrained or requantized. The file was made byscripts/mtp-graft.py
(gguf 0.19.0, numpy 2.5.3) from two inputs:
| Role | Repository @ revision | File | sha256 |
|---|---|---|---|
| Base (every tensor and key copied byte for byte) | HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive @ f12a584f |
Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf |
bbef58c3…4231 |
Donor (contributes only blk.40.*, the MTP head, and nextn_predict_layers) |
unsloth/Qwen3.6-35B-A3B-MTP-GGUF @ 5bc3e238 |
Qwen3.6-35B-A3B-UD-Q4_K_M.gguf |
0b21525e…c58b |
block_count goes from 40 to 41. Speculation never changes what the model says:
every drafted token is verified by the base model, so a head trained on the
original model costs only acceptance (0.79–0.86 measured on Apple Metal, about
0.6 on CUDA at 2 drafted tokens).
The graft is deterministic. Built independently on two machines, the output was
byte-identical:
sha256 8290c5bae3c744a9654281358e45669e158a9af2f0a23ad2a939bcb84528204f
bytes 21695616448
Use
llama-server -m Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_M-MTP.gguf \
--mmproj mmproj-Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-f16.gguf \
--spec-type draft-mtp --spec-draft-n-max 2 \
--temp 0.6 --top-p 0.95 --top-k 20 --min-p 0
The vision projector is not copied here; take it from
HauhauCS's repository.
Credits and license
- Original model: Qwen/Qwen3.6-35B-A3B by the Qwen team, Alibaba Cloud.
- Uncensored weights and quantization: HauhauCS.
- MTP head conversion: unsloth.
All three are released under the Apache License 2.0, and so is this file. The
license text is in LICENSE. "Qwen" names the model this is derived
from; it does not imply endorsement by the Qwen team.