This is HauhauCS/Qwen3.6-27B-Uncensored-HauhauCS-Balanced (Q4_K_P) with the Qwen 3.6 MTP "nextn" head grafted in, so it runs speculative decoding on llama.cpp out of the box — roughly 2× faster single-stream decode, with byte-identical output to the original (speculative decoding is lossless; the draft head only proposes tokens, the model verifies every one).

The upstream HauhauCS GGUFs ship without the MTP head, so there was no MTP-accelerated uncensored Qwen-27B available. This fills that gap.

What's different

	upstream Q4_K_P	this Q4_K_P-MTP
MTP / speculative decode	❌ none	✅ baked-in `nextn` head
Single-stream decode (4090, q8 KV, ctx 8K)	39.9 t/s	80.6 t/s (2.02×)
Draft acceptance	—	~64%
Output quality	identical	identical (lossless)
Vision (mmproj)	✅	✅

How it was made

The MTP head is an extra decoder block (blk.64.*, 15 tensors) plus two metadata keys (block_count → 65, nextn_predict_layers = 1). Those tensors were transplanted from unsloth/Qwen3.6-27B-MTP-GGUF (the only public source of the Qwen-27B MTP head) into the HauhauCS Q4_K_P file, with all other tensors/metadata copied faithfully. Because HauhauCS is a near-lossless abliteration of the same Qwen/Qwen3.6-27B base the head was trained on, draft acceptance stays high. Mixed per-tensor quant within a single GGUF is fully supported by llama.cpp.

Usage (llama.cpp)

bash

llama-server -m Qwen3.6-27B-Uncensored-HauhauCS-Balanced-Q4_K_P-MTP.gguf \
  --mmproj mmproj-Qwen3.6-27B-Uncensored-HauhauCS-Balanced-f16.gguf \
  -ngl 99 --flash-attn on -c 65536 \
  --spec-type draft-mtp --spec-draft-n-max 3 \
  -ctk q8_0 -ctv q8_0 --jinja

MTP: --spec-type draft-mtp (the head is in the main GGUF — no --model-draft needed). Requires a build with the merged Qwen MTP path (mainline ≥ b9542).
Context: fits ~65K tokens at q8 KV on a single 24 GB GPU with the vision projector loaded.
Chat template: the embedded template is stock-official Qwen 3.6. For agentic/tool-use, the unsloth fixed template (chat_template_unsloth.jinja, included) is recommended — it removes two over-eager exceptions and fixes tool-call argument serialization. Pass it with --chat-template-file chat_template_unsloth.jinja.
Thinking model: pass enable_thinking: false (template kwarg) for short/structured outputs; inline /no_think is not honored.

ik_llama.cpp (alternative)

ik_llama.cpp runs the same baked-in MTP head, but with its own flag dialect — -fa 1 instead of --flash-attn on, and -mtp --draft-max 3 instead of --spec-type draft-mtp (no --model-draft; the head is in the GGUF):

bash

./build/bin/llama-server \
  -m Qwen3.6-27B-Uncensored-HauhauCS-Balanced-Q4_K_P-MTP.gguf \
  --mmproj mmproj-Qwen3.6-27B-Uncensored-HauhauCS-Balanced-f16.gguf \
  -ngl 99 -fa 1 -c 65536 -ctk q8_0 -ctv q8_0 \
  -mtp --draft-max 3 \
  --jinja --host 127.0.0.1 --port 8080

This is a dense model, so MTP is a clear win on ik_llama too (~83 t/s measured on a 4090 for a comparable Qwen-27B-MTP build). Note: some ik_llama builds segfault on single-GPU + MTP — if so, run dual-GPU (-sm layer -ts 1,1) or fall back to mainline.

Files

Qwen3.6-27B-Uncensored-HauhauCS-Balanced-Q4_K_P-MTP.gguf — weights + grafted MTP head
mmproj-Qwen3.6-27B-Uncensored-HauhauCS-Balanced-f16.gguf — vision projector
chat_template_unsloth.jinja — recommended (bug-fixed) chat template

Credits & license

HauhauCS — the uncensored (abliterated) base weights.
unsloth — the Qwen 3.6 MTP head (donor) and the fixed chat template.
Qwen — the Qwen3.6-27B base model.

Apache-2.0, inherited from the upstream models. This is a redistribution with an added (lossless) MTP head — all model behavior and quality is HauhauCS's; only decode speed changes.

hyperhuzaifa/Qwen3.6-27B-Uncensored-MTP-GGUF

Qwen3.6-27B-Uncensored-HauhauCS-Balanced — Q4_K_P + MTP (GGUF)