---

license: apache-2.0

base_model:

- Qwen/Qwen3.6-35B-A3B

tags:

- gguf

- llama.cpp

- mtp

- speculative-decoding

- qwen3

- moe

- not-for-all-audiences

language:

- en

- zh

pipeline_tag: text-generation

---

Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive — MTP-grafted GGUF

> ⚠️ Not-for-all-audiences. This is an uncensored fine-tune of Qwen3 with reduced

> refusal behaviour. Use responsibly and in accordance with applicable law and the

> upstream licenses. The maintainers of the source models are not responsible for outputs.

A Q6_K GGUF of the Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive model, with a

Multi-Token-Prediction (MTP / nextn) head grafted on so that llama.cpp's

draft-mtp speculative decoding can be exercised.

What this is

The original Uncensored Q6_K_P GGUF shipped without the MTP head (the nextn

tensors were dropped during quantization). This file restores a single MTP layer by

transplanting blk.40 (20 tensors, including `nextn.eh_proj / enorm / hnorm /

shared_head_norm`) from a matching Qwen3.6-35B-A3B GGUF that retained it, and sets

block_count = 41, nextn_predict_layers = 1.

• Architecture: qwen35moe (35B total, ~3B active, 256 experts / 8 used)
• Quant: Q6_K (MTP eh_proj is Q8_0, norms F32)
• The transplant is byte-exact — no requantization of the base weights.

Performance — MTP gives ~+27% generation speedup

Measured with llama-bench on an Apple M3 Max (Metal), tg128, real prompt:

| Setting | Throughput | Draft acceptance | Mean accepted len |

|---|---|---|---|

| Baseline (no speculation) | 65.6 t/s | — | — |

| draft-mtp n_max=1 | 83.6 t/s | 90.2% | 1.90 tok/step |

| draft-mtp n_max=2 | 83.7 t/s | 83.8% | 2.68 tok/step |

| draft-mtp n_max=3 | 80.8 t/s | 76.5% | 3.29 tok/step |

| draft-mtp n_max=4 | 81.7 t/s | 75.2% | 4.01 tok/step |

The grafted MTP head reaches ~90% acceptance despite the backbone being an

uncensored fine-tune — i.e. the fine-tune diverges little from the base the MTP head was

trained on. Best operating point: --spec-draft-n-max 1 or 2, for roughly a +27%

generation throughput gain over plain autoregressive decoding.

> Note: acceptance rate is highly sensitive to having a real prompt/context. With an

> empty prompt the draft head has nothing to condition on and acceptance collapses (~31%);

> always benchmark with representative input.

Usage

# Plain generation
llama-cli -m Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q6_K_P-MTP.gguf -ngl 99 -c 8192

# MTP speculative decoding (requires a llama.cpp build with draft-mtp support)
llama-bench -m <file>.gguf -ngl 99 -n 64 --spec-type draft-mtp --spec-draft-n-max 2

Provenance & licenses

• Backbone: Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive (uncensored fine-tune of Qwen3).
• MTP layer: transplanted from a Qwen3.6-35B-A3B GGUF that retained the nextn head.
• Both derive from Qwen3 (Apache-2.0). Please honor the original Qwen license and the

respective fine-tuners'/quantizers' terms. If you are a rights holder and want

attribution corrected or content removed, open a discussion on this repo.