Modelpublic

barozp/Qwen3.6-28B-REAP20-A3B-GGUF

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

81likes6.2kdownloads

Model Card

Qwen3.6-28B-REAP20-A3B — GGUF Quantizations

GGUF quantizations of 0xSero/Qwen3.6-28B-REAP20-A3B, a 20% expert-pruned variant of Qwen/Qwen3.6-35B-A3B using the REAP (Router-weighted Expert Activation Pruning) method.

Available Files

File	Quant	Size	BPW	Description
`Qwen3.6-28B-REAP20-A3B-BF16.gguf`	BF16	~56.5 GB	16.0	Full precision, for re-quantization
`Qwen3.6-28B-REAP20-A3B-Q8_0.gguf`	Q8_0	~30 GB	8.0	Near-lossless, large file
`Qwen3.6-28B-REAP20-A3B-Q6_K.gguf`	Q6_K	~23 GB	6.56	Near-lossless, recommended for high quality
`Qwen3.6-28B-REAP20-A3B-Q5_K_M.gguf`	Q5KM	~20 GB	5.68	High quality, larger size
`Qwen3.6-28B-REAP20-A3B-Q5_K_S.gguf`	Q5KS	~19 GB	5.52	High quality, slightly smaller
`Qwen3.6-28B-REAP20-A3B-Q4_K_M.gguf`	Q4KM	~17 GB	4.89	Recommended — best quality/size balance
`Qwen3.6-28B-REAP20-A3B-Q4_K_S.gguf`	Q4KS	~16 GB	4.63	4-bit small
`Qwen3.6-28B-REAP20-A3B-Q3_K_L.gguf`	Q3KL	~15 GB	4.27	3-bit large
`Qwen3.6-28B-REAP20-A3B-Q3_K_M.gguf`	Q3KM	~14 GB	3.91	3-bit medium
`Qwen3.6-28B-REAP20-A3B-Q3_K_S.gguf`	Q3KS	~13 GB	3.66	3-bit small
`Qwen3.6-28B-REAP20-A3B-IQ3_XXS.gguf`	IQ3_XXS	~12 GB	3.06	Ultra-small, imatrix-based
`Qwen3.6-28B-REAP20-A3B-Q2_K.gguf`	Q2_K	~11 GB	2.96	Smallest size, lowest quality

Model Details

Property	Value
Architecture	Qwen3.6 MoE (hybrid Gated DeltaNet + MoE)
Parameters	~28B total / ~3B active per token
Experts	205 total / 8 active per token (pruned from 256)
Context Length	262,144 tokens
Original dtype	BF16
Quantization source	BF16 GGUF from 0xSero/Qwen3.6-28B-REAP20-A3B-GGUF
Quantization tool	llama.cpp
imatrix	Used for IQ3_XXS (from source repo)
License	Apache 2.0

Quantization Process

bash

# 1. Download BF16 GGUF from source
huggingface-cli download 0xSero/Qwen3.6-28B-REAP20-A3B-GGUF \
  --include "model.bf16.gguf" --local-dir ./

# 2. Download imatrix (for IQ quants)
huggingface-cli download 0xSero/Qwen3.6-28B-REAP20-A3B-GGUF \
  --include "imatrix.dat" --local-dir ./

# 3. Quantize (example: Q4_K_M)
llama-quantize model.bf16.gguf Qwen3.6-28B-REAP20-A3B-Q4_K_M.gguf Q4_K_M

# 4. Quantize with imatrix (example: IQ3_XXS)
llama-quantize --imatrix imatrix.dat model.bf16.gguf \
  Qwen3.6-28B-REAP20-A3B-IQ3_XXS.gguf IQ3_XXS

Usage

llama.cpp

bash

llama-cli \
  -m Qwen3.6-28B-REAP20-A3B-Q4_K_M.gguf \
  -ngl 99 -c 4096 \
  -p "Your prompt here"

llama-server (OpenAI-compatible API)

bash

llama-server \
  -m Qwen3.6-28B-REAP20-A3B-Q4_K_M.gguf \
  -ngl 99 -c 4096 \
  --port 8080

LM Studio / Jan / Ollama

Download the .gguf file and load it directly in your preferred local inference UI.

Hardware Requirements

Config	VRAM / RAM
Full GPU (Q4KM, recommended)	20+ GB VRAM
Hybrid CPU+GPU (Q4KM)	10 GB VRAM + 10 GB RAM
CPU only (Q4KM)	24+ GB RAM

About the Original Model

0xSero/Qwen3.6-28B-REAP20-A3B applies REAP expert pruning (arXiv:2510.13999) to remove 20% of MoE experts (51 of 256 per layer) from Qwen3.6-35B-A3B, while preserving routing behavior via router weight renormalization. Active parameters per token remain unchanged at ~3B. The result is a ~25% smaller model with competitive generation quality across coding, reasoning, and knowledge benchmarks.

License

Apache 2.0 — see Qwen License.