---

license: apache-2.0

base_model: Qwen/Qwen3.6-27B

tags:

- qwen

- qwen3.6

- 27b

- mtp

- nextn

- speculative-decoding

- gguf

- quantized

- imatrix

pipeline_tag: text-generation

---

Qwen3.6-27B-MTP-IMAT-IQ4_XS-Q8nextn (GGUF)

Dense 27B Qwen3.6 trunk + embedded NextN-MTP head, quantized for single-GPU inference.

• Trunk: IQ4_XS (imatrix-calibrated)
• MTP head: Q8_0 (NextN, kv_only_nextn=true)
• File size: ~14.5 GB
• Min VRAM: ~17 GB at 8K ctx, ~22 GB at 96K+ ctx (KV at Q8/Q8)

The MTP head is embedded inside the GGUF — no separate drafter file. Recent llama.cpp builds activate it via --spec-type draft-mtp.

Quick start (no spec decode)

llama-server -m Qwen3.6-27B-MTP-IMAT-IQ4_XS-Q8nextn.gguf \
  -ngl 999 -fa on -c 32768 --parallel 1 \
  -ctk q8_0 -ctv q8_0 --kv-unified \
  --host 0.0.0.0 --port 8080 --jinja

With speculative decoding (recommended)

Requires llama.cpp built from master after ggml-org/llama.cpp#22673:

llama-server -m Qwen3.6-27B-MTP-IMAT-IQ4_XS-Q8nextn.gguf \
  -ngl 999 -fa on -c 32768 --parallel 1 \
  -ctk q8_0 -ctv q8_0 --kv-unified \
  --spec-type draft-mtp \
  --spec-draft-n-max 4 \
  --spec-draft-p-min 0.75 \
  --host 0.0.0.0 --port 8080 --jinja

--spec-type draft-mtp binds the embedded NextN MTP head as a drafter via LLAMA_CONTEXT_TYPE_MTP. --spec-draft-n-max 4 is max draft chain length per round. --spec-draft-p-min 0.75 is essential — without it the chain drafts low-confidence tokens that the target rejects, halving the speedup.

Benchmarks

Single-stream decode on a 24 GB consumer GPU (RTX 3090 Ti), CUDA build, full-GPU offload, FA on, N=512 generated tokens, 5-warm-run mean.

Production config (ctx=192K, KV Q8/Q8)

| config | decode tok/s | accept rate |

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

| no spec (greedy) | ~48 | – |

| --spec-type draft-mtp --spec-draft-n-max 3 --spec-draft-p-min 0.75 | ~73 (+52%) | ~70% |

Smaller ctx variant (ctx=8192, KV q4_0/q4_0)

| config | tok/s |

|---|---|

| greedy | 47.7 |

| --spec-draft-n-max 4 --spec-draft-p-min 0.0 | 78.7 |

| --spec-draft-n-max 3 --spec-draft-p-min 0.75 | 77.7 |

Dense 27B benefits from p_min even on q4 — sampler cost is amortized over the deeper accept gains. MoE variants (35B-A3B) prefer p_min=0 because per-step softmax overhead outweighs the chain-termination savings.

Quant details

• Trunk weights: IQ4_XS calibrated with an imatrix derived from a mixed-domain calibration set (code + multilingual prose).
• MTP head: kept at Q8_0 (NextN). Draft quality is sensitive to head precision — quantizing the head further significantly drops accept rate.
• KV cache type: Q8_0/Q8_0 runs cleanly. Lower precision on V saves memory but trims accept rate a few pp.

llama.cpp requirements

• Build supporting Qwen3.5 / Qwen3.6 NextN MTP (LLM_ARCH_QWEN35_MTP / LLM_ARCH_QWEN3MOE_MTP). Merged upstream in #22673.
• For zero-config --spec-type draft-mtp (no separate drafter file), the server must wire LLAMA_CONTEXT_TYPE_MTP inline-MTP context init — on master since the PR landed.
• --spec-draft-p-min 0.75 requires recent llama.cpp that honors the p_min parameter on the DRAFT_MTP path. If your build doesn't, set p_min to 0 — you'll still get speedup, just less.

Sister models

• Without imatrix: localweights/Qwen3.6-27B-MTP-IQ4_XS-Q8nextn-GGUF
• 35B-A3B MoE variant: localweights/Qwen3.6-35B-A3B-MTP-IMAT-IQ4_XS-Q8nextn-GGUF

Reference

• Base model: <https://huggingface.co/Qwen/Qwen3.6-27B>
• MTP architecture: NextN-style embedded prediction layer, kv_only_nextn=true.

Apache 2.0 license.