license: apache-2.0
base_model:

• huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated
• Jackrong/Qwopus3.6-35B-A3B-Coder-MTP-GGUF
  base_model_relation: quantized
  quantized_by: xero0000
  pipeline_tag: text-generation
  library_name: gguf
  tags:
• gguf
• qwen35moe
• moe
• mixed-precision
• imatrix
• ik_llama.cpp
• mtp
• speculative-decoding
• q2_k
• uncensored
• abliterated

🕴️ G-Man — Huihui Qwen3.6-35B-A3B abliterated Mixed q2_K + Transplanted MTP Head

Black Mesa mixed-quant series · operates outside the rules (uncensored) — and now arrives sooner.

The G-Man (plain mixed quant) mixed-precision GGUF of Huihui Qwen3.6-35B-A3B abliterated, with one
addition: the multi-token-prediction (MTP) head from
Qwopus-3.6-Coder
surgically grafted on, enabling self-speculative decoding in
ik_llama.cpp.

TL;DR: identical outputs to the plain mixed quant, but ~87 tok/s on code and
~83 tok/s on prose instead of ~78, on an 18 GB dual-GPU desktop. Free speed,
verified token-by-token.

The head transplant

Huihui Qwen3.6-35B-A3B abliterated ships without an MTP head — but Qwopus-Coder is a fine-tune of the same
Qwen3.6-35B-A3B base, and its blk.40 nextn head (a full extra layer: attention + MoE +
eh_proj/enorm/hnorm glue, ~0.55 GB) predicts the next-next token from hidden states
that this model's hidden space is a close sibling of. So we graft it:

1. append the donor's 20 blk.40.* tensors after the target's 40 layers,
2. bump block_count 40 → 41,
3. set qwen35moe.nextn_predict_layers = 1.

Why this is safe: speculative decoding verifies every drafted token against this
model. A foreign head can never change the output distribution — a bad match only lowers
the acceptance rate (= less speedup). Measured across the series, acceptance tracks
fine-tune distance from the donor: base Qwen3.6 93/90 (code/prose t/s) > abliterated 87/83

Ornith 83/80 > AgentWorld 82/79, all against a 78 t/s no-MTP baseline.

Recipe

• Quant layout (same as the parent mixed quant): ffn_*_exps on blocks 13–26 →
  Q2_K with importance matrix (the CPU-offloaded set); everything else Q4_K;
  output-class Q6_K. ~4.9 bpw effective, ~18.7 GB.
• Head: blk.40 nextn/MTP layer from Qwopus-Coder mixed-q2k (Q4_K experts), grafted
  byte-exact.
• The mixed layout exists because decode on CPU-offload rigs is RAM-bandwidth-bound: only
  the offloaded-layer bytes matter, so those get Q2_K while GPU-resident tensors keep
  Q4_K quality.

Measured performance

Rig: RTX 3060 Ti 8 GB + RTX 3080 10 GB, DDR4, ik_llama.cpp, 128K ctx, greedy.

How to run

Requires ik_llama.cpp — its -mtp flag is
what drives the nextn head (mainline llama.cpp loads the file but ignores the head).

./llama-server -m Qwen3.6-35B-A3B-abliterated-mixed-q2k-MTP.gguf \
  --jinja --cache-type-k q4_0 --cache-type-v q4_0 --flash-attn on \
  --ctx-size 131072 --parallel 1 --n-gpu-layers 99 --ctx-checkpoints 8 \
  -ot 'blk\.(1[3-9]|2[0-9])\.ffn_(up|gate|down)_exps\.weight=CPU' \
  --tensor-split 44,56 --ubatch-size 256 \
  -mtp --ctx-size-draft 8192 \
  --no-mmap --threads 8 --no-warmup

Notes for 18 GB-class rigs:

• The MTP draft context costs VRAM, which is why this profile runs 128K ctx (not 256K),
  a small 8K draft context, and pins expert layers 13–29 to CPU (three more than the
  plain mixed profile). With less freed VRAM, flash-attention temp allocations OOM
  mid-decode. More VRAM → pin fewer layers and/or raise ctx.
• --ctx-checkpoints 8 caps ik's dynamically allocated SSM checkpoints (default 32 ×
  64 MiB ≈ 2 GB at deep context — an OOM trap on long agent sessions).
• Add --reasoning off --reasoning-budget 0 for tool/browser loops (drop them for deep chat).
• Drop -mtp (and the diet) and it behaves exactly like the parent mixed quant at 256K.

Credits

• Base model: huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated (Apache-2.0)
• MTP head donor: Jackrong/Qwopus3.6-35B-A3B-Coder-MTP-GGUF
• Mixed quant, imatrix, transplant & profiling: xero0000
• Series: Gordon (base) · Kleiner (coder) · G-Man (uncensored) · Vortigaunt (reasoner) · Alyx (agentic)