This is a "Hybrid" IQ4_KS GGUF quantization of Qwen3.6-27B that preserves the MTP (Multi-Token Prediction) layers, allowing for significantly faster text generation via speculative decoding.

Standard GGUF conversions often strip out MTP tensors to save a tiny bit of space. This model was carefully requantized from Radamanthys11/Qwen3.6-27B-MTP-Q8_0-GGUF using ik_llama.cpp to retain the MTP head while shrinking the VRAM requirements down to a highly efficient Q4 footprint.

Inference Speed (MTP vs Baseline)

Tested using llama-server on a 3090 to measure evaluation speed. Enabling Multi-Token Prediction with 1 draft token yields roughly a 16% speedup over standard inference.

Configuration	Speed	Setup / Flags
MTP Max 1	28.99 t/s	`-mtp --draft-max 1 --draft-p-min 0.0`
Baseline (No MTP)	24.99 t/s	(No MTP flags)
MTP Max 2	24.84 t/s	`-mtp --draft-max 2 --draft-p-min 0.0`

Perplexity

Measured against wiki.test.raw with n_ctx=512 over 580 chunks. Tests confirm that enabling MTP decoding does not negatively impact the perplexity score:

Quant	Size	Without MTP PPL	With MTP PPL
Hybrid IQ4_KS	16.8 GiB	6.9424 +/- 0.04574	6.9424 +/- 0.04574
Q4_K_M	15.7 GiB	7.0291 +/- 0.04648	7.0291 +/- 0.04648

Note: Lower is better.

Quantization Recipe

The custom ruleset used for the mixed-precision tensor overrides:

bash

custom="
# SSM State Logic
blk\..*\.ssm_alpha\.weight=f32
blk\..*\.ssm_beta\.weight=f32
blk\..*\.ssm_out\.weight=q8_0

# 1. Non-linear mapping strictly for attention
blk\..*\.attn_.*\.weight=iq4_nl

# 2. Sandwich boost (First 8 / Last 8) -> iq5_ks for ALL FFN tensors
blk\.[0-7]\.ffn_.*\.weight=iq5_ks
blk\.(5[6-9]|6[0-3])\.ffn_.*\.weight=iq5_ks

# 3. Global bottleneck boost -> iq5_ks for remaining ffn_down
blk\..*\.ffn_down\.weight=iq5_ks

# 4. Fallback -> iq4_ks for remaining ffn_gate / ffn_up
blk\..*\.ffn_.*\.weight=iq4_ks

# 5. High precision anchors
token_embd\.weight=q8_0
output\.weight=q8_0
"

custom=$(
  echo "$custom" | grep -v '^#' | \
  sed -Ez 's:\n+:,:g;s:,$::;s:^,::'
)

How This Was Made (Reproduction Steps)

This model was re-quantized directly from a Q8_0 intermediate model using --allow-requantize.

1. Generate the Imatrix from Q8_0 Note: GGML_CUDA_NO_PINNED=1 is used to prevent system RAM exhaustion on 24GB VRAM setups.

bash

GGML_CUDA_NO_PINNED=1 ./ik_llama.cpp/build/bin/llama-imatrix \
  -m ./Qwen3.6-27B-MTP-Q8_0.gguf \
  -f /path/to/ubergarm-imatrix-calibration-corpus-v02.txt \
  -o Qwen3.6-27B-MTP-imatrix.dat \
  --ctx-size 512 \
  -t 16 \
  --fit

2. Requantize Q8_0 directly to Hybrid IQ4_KS

bash

./ik_llama.cpp/build/bin/llama-quantize \
  --allow-requantize \
  --imatrix ./Qwen3.6-27B-MTP-imatrix.dat \
  --custom-q "$custom" \
  ./Qwen3.6-27B-MTP-Q8_0.gguf \
  ./Qwen3.6-27B-MTP-IQ4_KS.gguf \
  IQ4_KS 16

3. Test Perplexity

bash

wget https://huggingface.co/datasets/ikawrakow/validation-datasets-for-llama.cpp/resolve/main/wiki.test.raw.gz
gunzip wiki.test.raw.gz

# Standard Perplexity
./ik_llama.cpp/build/bin/llama-perplexity \
  -m ./Qwen3.6-27B-MTP-IQ4_KS.gguf \
  -f ./wiki.test.raw \
  -c 512 \
  -ngl 99

# Perplexity with MTP enabled
./ik_llama.cpp/build/bin/llama-perplexity \
  -m ./Qwen3.6-27B-MTP-IQ4_KS.gguf \
  -f ./wiki.test.raw \
  -c 512 \
  -ngl 99 \
  -mtp --draft-max 1 --draft-p-min 0.0

Quick Start Inference

Requires ik_llama.cpp. Be sure to pass the MTP flags (-mtp and --draft-max 1) to utilize the Multi-Token Prediction speedups!

bash

./ik_llama.cpp/build/bin/llama-server \
  -m ./Qwen3.6-27B-MTP-IQ4_KS.gguf \
  -c 10000 \
  -ngl 99 \
  -mtp --draft-max 1 --draft-p-min 0.0

RDson/Qwen3.6-27B-MTP-IQ4_KS-GGUF

Qwen3.6-27B-MTP Hybrid IQ4_KS GGUF

Inference Speed (MTP vs Baseline)

Perplexity

Quantization Recipe

How This Was Made (Reproduction Steps)

Quick Start Inference

RDson/
Qwen3.6-27B-MTP-IQ4_KS-GGUF