akqmffl/qwen3.6-27b-Q8_0-MTP-Claude-Opus-Reasoning-uncensored-GGUF

Qwen3.6 27B Q8_0 Native MTP Claude-Opus-Reasoning Uncensored GGUF

Important runtime requirement: to reproduce the benchmark performance for this GGUF, download and configure akqmffl/llama.cpp-lemonade-Runtime-for-qwen-3.6-MTP. Benchmark settings, release downloads, launch scripts, Lemonade integration, runtime environment variables, and supported hardware notes are maintained in that GitHub repository. Visit the GitHub README and release notes for the current setup procedure and benchmark configuration.

This repository contains a Q8_0 GGUF build of qwen3.6-27b-Q8_0-MTP-Claude-Opus-Reasoning-uncensored.gguf. It is intended for the custom llama.cpp / Lemonade ROCm runtime above, especially the Windows AMD Ryzen AI Max Series APU / RDNA 3.5 / gfx1151 path documented there.

Preserved MTPs

This GGUF preserves the Native MTP tensor set. The model file was checked at the GGUF tensor table level and contains 15 MTP-related tensors:

1. blk.64.attn_k.weight
2. blk.64.attn_k_norm.weight
3. blk.64.attn_norm.weight
4. blk.64.attn_output.weight
5. blk.64.attn_q.weight
6. blk.64.attn_q_norm.weight
7. blk.64.attn_v.weight
8. blk.64.ffn_down.weight
9. blk.64.ffn_gate.weight
10. blk.64.ffn_up.weight
11. blk.64.nextn.eh_proj.weight
12. blk.64.nextn.enorm.weight
13. blk.64.nextn.hnorm.weight
14. blk.64.nextn.shared_head_norm.weight
15. blk.64.post_attention_norm.weight

The MTP component was trained and applied using target-model-aligned MTP distillation against the qwen3.6-27b-Claude-Opus-Reasoning-uncensored target model. In other words, the MTP path is intended to follow the target model's next-token distribution for speculative decoding rather than behaving as an unrelated draft model.

Model Overview

Item	Value
Format	GGUF v3
Quantization	Q8_0
File	qwen3.6-27b-Q8_0-MTP-Claude-Opus-Reasoning-uncensored.gguf
File size	27.05 GiB
SHA-256	BB0A4A050D598FD88D570B326B8331B9B11767ACB10E9CC8B0B899F4C3350539
Tensor count	866
Metadata KV count	40
Native MTP tensor count	15
Target alignment	Target-model-aligned MTP distillation
Intended runtime	Custom llama.cpp / Lemonade ROCm runtime for gfx1151

Abliteration Parameters

This GGUF is packaged from the qwen3.6-27b-Claude-Opus-Reasoning-uncensored target-model line. This model card does not claim a new ablation pass or publish separate ablation parameter values for this GGUF. The important alignment note for this upload is that the Native MTP component was trained against that uncensored target model by target-model-aligned MTP distillation before being applied to the Q8_0 GGUF.

Performance

The performance-sensitive path depends on both the GGUF and the custom runtime. Use the GitHub package for exact setup and benchmark reproduction:

• Runtime and setup: akqmffl/llama.cpp-lemonade-Runtime-for-qwen-3.6-MTP
• Release page: b1294

Recorded runtime-package anchors:

Lane	Shape	Recorded result
n=2 production	128K context, 4096 generated tokens, FA off	15.131 tok/s, 92.286% acceptance
n=2 reasoning off	128K context, 4096 generated tokens, FA off	15.413 tok/s, 93.182% acceptance
n=3 chat stream	8192 context, 100 generated tokens, FA off	18.050 tok/s, 97.333% acceptance, exact at 100 tokens

AMD Ryzen AI Max / RDNA 3.5 / gfx1151 Q8 Baselines

The comparison set below is restricted to Qwen3.6-27B-MTP-Q8 and Qwen3.6-27B-MTP-UD-Q8_K_XL public rows on AMD Ryzen AI Max / Strix Halo / Radeon 8060S / RDNA 3.5 / gfx1151 class hardware that report both generation speed and MTP draft acceptance. Non-Q8 rows are intentionally excluded from this table.

Source	Hardware / backend	Model and runtime	Workload	tok/s	MTP acceptance	Comparison note
This upload, short stream	AMD Ryzen AI Max Series APU / RDNA 3.5 / gfx1151, custom ROCm runtime	Qwen3.6 27B Q8_0 Native MTP, n=3, FA off	8192 context, 100 generated tokens	18.050	97.333%, exact at 100 tokens	Release anchor for short interactive generation.
This upload, long output	AMD Ryzen AI Max Series APU / RDNA 3.5 / gfx1151, custom ROCm runtime	Qwen3.6 27B Q8_0 Native MTP, n=2, FA off	128K context, 4096 generated tokens	15.413	93.182%	Release anchor for long 128K-context output.
Reddit AMD ROCm Windows PR 22673 report	Ryzen AI Max+ 395, Radeon 8060S iGPU, RDNA 3.5 gfx1151, Windows 11 ROCm	Qwen3.6-27B-MTP-UD-Q8_K_XL.gguf, PR 22673 llama.cpp build	128K context, q8_0 KV, thinking on, draft-MTP	12.13 tokens/sec	64-69%	Public same-hardware UD-Q8_K_XL row with both speed and acceptance.
Qiita AMD Ryzen AI MAX+ 395 Qwen3.6 MTP Q8_0 run	AMD Ryzen AI MAX+ 395 / gfx1151, llama.cpp ROCm	Qwen3.6-27B-MTP-Q8_0.gguf, draft-MTP	8192 context, 142 output tokens, temperature 0.7, seed 42	10.15 tokens/sec	60/243 = 24.7%	Public Q8_0 same-hardware row; acceptance is much lower than this release anchor.

A static HTML comparison view is included in assets/amd-gfx1151-benchmark-comparison.html.

These anchors are hardware- and runtime-specific. A generic llama.cpp build may load the GGUF, but it should not be expected to match the recorded MTP throughput without the runtime package and settings documented in the GitHub repository.

Files

File	Description
qwen3.6-27b-Q8_0-MTP-Claude-Opus-Reasoning-uncensored.gguf	Q8_0 GGUF with Native MTP preserved and target-model-aligned MTP distillation applied

Usage

1. Download the GGUF from this repository.
2. Download the matching runtime package from akqmffl/llama.cpp-lemonade-Runtime-for-qwen-3.6-MTP.
3. Follow the GitHub README for standalone llama.cpp or Lemonade setup.
4. Use this GGUF path as the model path in the GitHub package's launch scripts or Lemonade configuration.

Reference

This model-card structure was prepared after reviewing the Hugging Face GGUF reference repository llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved-NVFP4-GGUF. Only the structure was used; benchmark/quality metrics from that repository are not copied into this card.

License

Apache-2.0 metadata is declared to match the referenced Qwen/GGUF lineage style. Users remain responsible for verifying upstream model licensing, local deployment requirements, and acceptable-use constraints for their own environment.